How Google AI Mode and AI Overviews Select Sources

How Google AI Mode and AI Overviews Select Sources

Google's AI search experiences, AI Mode and AI Overviews, select sources using fundamentally different criteria from traditional organic rankings. Understanding each one is the foundation of any AI visibility strategy in 2026.

Key takeaways:

• Only 14% of URLs cited in AI Mode also rank in the top 10 of traditional Google search results
• AI Overviews now appear in approximately 48% of all tracked queries, up from 30% a year ago
• 62% of AI Overview citations come from pages outside the organic top 10 as of early 2026
• AI Mode uses a query fan-out technique that selects sources at a granular level traditional SEO never needed to address

Most of the B2B software brands we work with at FirstMotion assume their Google rankings carry over to AI Mode and AI Overviews. The data says otherwise. The gap between organic and AI is now large enough to demand a separate strategy, and this guide explains exactly what drives citation selection on each platform.

What is Google AI Mode and how does it work in Google Search?

Google AI Mode is a dedicated tab within Google Search powered by Gemini 2.5. It generates synthesised, conversational responses to complex queries using more advanced reasoning than traditional search, interprets queries through text, images, or voice, and retrieves real-time information from the live web rather than a static index.

AI Mode reduces the need to reformulate searches and visit multiple websites, because it handles multi-part questions and performs multiple background searches simultaneously. Google confirmed in its August 2025 announcement that AI Mode now reaches 180 countries and territories in English, making it the most powerful AI search experience Google has ever deployed globally.

AI Mode also uses multimodal capabilities that go beyond text. Through Search Live, it lets users point their camera at real-world objects and ask questions about what they see, using computer vision to analyse environments in real time. Google's Agentic Vision within Gemini 3 Flash takes this further, using computer vision to improve image recognition accuracy, automating visual analysis tasks that previously required manual processes and delivering superior accuracy compared to manual inspection methods.

What are Google AI Overviews and how do AI Overview citations work?

Google AI Overviews is a separate product from AI Mode. It appears directly on the main Google search results page as an AI generated summary above traditional organic listings, without any tab switch required. AI Overviews launched officially on May 14, 2024, focusing specifically on increasing visibility in AI generated search summaries for informational queries.

A critical distinction: AI Overviews cite passages, not entire pages. The citation unit is a specific extractable answer within a page, not the overall authority of the domain. BrightEdge's year-over-year analysis confirms AI Overviews now trigger on approximately 48% of tracked queries, up from 30% a year ago.

For local businesses, this prevalence matters significantly. Queries about local services, healthcare providers, and professional services increasingly surface AI Overviews rather than traditional organic listings. Brands that earn an AI Overview citation see a 35% increase in organic clicks compared to competitors that don't appear in the overview, according to Seer Interactive's analysis of queries across 42 organisations.

How AI Mode works: query fan-out and AI generated responses

AI Mode's source selection starts before it retrieves a single page. Ahrefs' AI Mode guide confirms that AI Mode uses a query fan-out technique that takes the original query, divides it into multiple sub-queries, and sends each to Google's index independently. A single question in the AI Mode search bar can trigger dozens of parallel searches across different facets of the same topic.

This architecture produces a very different AI response from what traditional search generates. A page ranking position one for the primary query can lose citation slots to candidate pages that answer sub-queries well, even when those exact URLs don't rank for the original question. AI Mode queries tend to be significantly longer and more conversational than traditional search queries, which means AI Mode selects content at a much more granular and intent-specific level.

AI Mode also provides a more detailed analysis of complex topics than any AI generated answer or featured snippet. When users want to dive deeper, they ask follow-up questions within the same session, and AI Mode performs additional query fan-out rounds to retrieve more specific context. This extended session behaviour means multiple brands can earn citations across a single conversation, creating citation opportunities that don't exist in any other Google search format.

How AI Mode selects sources: what the data shows

SE Ranking's August 2025 study analysed AI Mode responses across a large keyword set and produced three findings that fundamentally change how AI visibility needs to be measured.

The 14% URL overlap is the most strategically significant figure. It confirms that AI Mode rarely references the pages Google ranks highest in traditional search results, and operates on a fundamentally different approach to content relevance. For brands tracking AI visibility through organic rankings alone, these figures confirm that organic search results are almost entirely missing what AI Mode actually does with their content. User feedback signals, including follow-up question patterns and session dwell time, also influence which specific pages get selected over time.

Google's AI Overviews: AI Overview visibility data and citation patterns

AI Overviews and AI Mode share the same Google infrastructure but select sources differently. AI Overviews focus on informational queries, cite passages rather than entire pages, and correlate more strongly with organic rankings than AI Mode, though that correlation has weakened significantly in 2026.

Digital Applied's post-I/O 2026 analysis shows that in July 2025, 76% of AI Overview citations came from pages ranking in the organic top 10. By March 2026, that figure had fallen to 38%, a 50% relative decline in eight months. Ahrefs' March 2026 analysis confirms that 62% of AI Overview citations now come from pages outside the top 10 organic results, as top-10 citation rates fell from 76% to 38% in eight months.

AI Overviews also push traditional organic listings further down the page. The average overview now exceeds 1,200 pixels in height, displacing organic search results, blue links, and featured snippets significantly below the fold on AI Overview-triggered queries. Ahrefs' updated December 2025 study found that the presence of an AI Overview now correlates with a 58% lower average clickthrough rate for the top-ranking page, updated from their initial 34.5% finding in April 2025.

Generative AI in Google Search: AI Mode vs AI Overviews vs traditional search

The clearest way to understand how generative AI has changed source selection is to compare all three surfaces directly. Each operates on different signals, rewards different content properties, and delivers a different user experience.

The most important distinction is the citation unit. AI Overviews cite passages; AI Mode selects at the sub-query level. Both systems evaluate specific content within a page, not the overall authority of the page itself. That's why candidate pages outside the top 10 regularly earn AI citations when they contain the most directly answerable passage for a specific sub-topic.

How AI Overview visibility differs from AI Mode visibility

AI Mode visibility and AI Overview visibility are distinct metrics that require separate tracking strategies. Ahrefs' analysis of 540,000 query pairs found that AI Mode and AI Overviews cite the same URLs only 13.7% of the time. A brand can earn strong AI Overview citations without appearing in AI Mode responses at all, and vice versa.

AI Overview visibility aligns more closely with traditional organic rankings, topical authority, and content quality. Pages that rank well for informational queries, carry schema markup including Article schema and HowTo schema, and cover topics with genuine contextual understanding earn AI Overview citations at higher rates. AIO focuses specifically on synthesising helpful links and cited pages for the user's initial query, meaning content that directly and clearly answers common questions performs best.

AI Mode visibility requires a different approach because of the query fan-out architecture. AI Mode visibility depends on covering the full range of sub-topics a complex query generates, not just the primary keyword. A brand that answers one aspect of a query well but leaves adjacent sub-queries uncovered will see inconsistent AI Mode citation patterns, regardless of domain authority or traditional search results performance.

What drives AI Overview citations and AI generated answers across both platforms

Both AI Mode and AI Overviews reward the same underlying content properties, though they weight them differently. These signals consistently improve citation likelihood across both platforms:

• Direct answers first: content that answers the specific query in the opening paragraph gets extracted more reliably. An AI generated answer draws from the most immediately relevant passage, not the most comprehensive page
• Topical depth: covering all the sub-topics a query fan-out generates means more sub-queries find a citable passage within the same domain, keeping multiple brands from occupying citation slots your content should fill
• Schema markup: Article schema, HowTo schema, and FAQ schema all improve passage-level extractability for specific pages. Google Search Central confirms JSON-LD is the recommended implementation
• Content freshness: AI systems favour recently updated content with current statistics and contemporary references on cited pages
• Entity clarity: naming the brand, topic, and use case explicitly in titles, headings, and opening paragraphs helps Google's AI systems anchor AI citations accurately
• Technical SEO foundations: pages that load quickly and render correctly for AI crawlers pass eligibility requirements before any relevance evaluation begins
• Topical authority: domains that cover a topic area comprehensively build the citation trust AI Mode's query fan-out needs to return to the same domain repeatedly across multiple searches

Content quality has become the dividing line between brands that appear consistently in AI generated answers and brands that don't. AI systems automate relevance evaluation at scale, delivering superior accuracy compared to any manual content audit process.

How AI Mode personalisation affects source selection and where brand appears

AI Mode's personalisation layer adds a dimension to source selection with no direct equivalent in traditional SEO or AI Overview optimisation. When users opt in, AI Mode references past searches, location data, and activity from the Google app and Google Maps to generate an AI powered response tailored to their personal context.

The same query from two different users can produce entirely different cited sources and different AI response content. Content that speaks to specific use cases, buyer stages, and geographic contexts, including local businesses and region-specific solutions, earns more citations in personalised responses for those segments. A brand that only publishes generic category-level content won't appear in personalised AI Mode responses, even when it ranks well in traditional organic search results.

For B2B brands, topical depth across the full buyer journey is essential. AI Mode needs enough relevant content across an entire topic area to construct personalised responses. Brands that publish at multiple depth levels, from overview articles to detailed technical guides, give AI Mode more citation options across different user contexts.

How to measure AI generated visibility and AI Mode citations

Measuring AI visibility requires different tools from traditional rank tracking. Organic rankings are a necessary but insufficient proxy for AI citation performance, and the gap between the two continues to widen across all search engines incorporating generative AI.

Platforms that now track AI visibility directly include:

• SE Ranking AI Search Toolkit: tracks AI Overview citations and AI Mode citations at keyword level, with volatility monitoring across multiple searches of the same query
• Ahrefs Brand Radar: indexes AI Mode responses and lets brands check citation frequency for exact URLs across a growing query dataset
• BrightEdge Generative Parser: monitors AI Overview presence and overview citations with year-over-year trends across industry verticals
• Semrush AI Toolkit: tracks AI Overview visibility alongside traditional organic results for comparison across SEO platforms

FirstMotion's AI search audit starts by mapping a brand's citation footprint across AI Mode and AI Overviews, comparing it against competitor citation rates, and identifying the specific content and technical gaps that explain the difference. Continuous monitoring of AI citation rates is the only reliable signal of AI search performance because organic visibility no longer predicts it.

Featured snippets, blue links, and what AI search replaces

AI Mode and AI Overviews don't just complement traditional search. For informational queries, they're actively replacing featured snippets and blue links as the primary way users receive answers. Understanding this displacement helps brands prioritise where to focus their SEO strategy and content investment.

Featured snippets were the first step in Google's transition from returning links to returning answers directly. AI Overviews took that further by synthesising answers from multiple sources. AI Mode goes further still, replacing the entire traditional search results experience with a conversational AI response that handles the full research session without requiring multiple clicks to individual websites.

The brands earning consistent AI citations treat this as a content architecture problem, not a keyword problem. Topical depth, structured data coverage, and passage-level clarity determine AI citation outcomes. The AI search revolution in B2B SaaS has already made these signals the primary competitive differentiator in organic visibility for informational queries.

If your brand isn't appearing in AI Mode or AI Overviews, here's where to start

Most of the brands we audit at FirstMotion aren't invisible in AI search because their content is low quality. They're invisible because their content strategy was built for a different citation system. A targeted audit of citation gaps, schema markup coverage, and topical depth usually reveals fixable issues within the first session.

If you want to understand exactly why your brand isn't being cited and what to prioritise first, talk to the FirstMotion team. We'll map your AI citation footprint and show you the fastest path to AI search visibility.

Frequently Asked Questions

What is the difference between AI Mode and AI Overviews?

AI Mode is a dedicated tab within Google Search that generates conversational, multi-part answers to complex queries using Gemini, with follow-up question capability and deep personalisation. AI Overviews appears on the main Google search results page as an AI generated answer above organic results, focusing on informational queries. Both cite sources but use different selection criteria and share only 13.7% URL overlap.

How does AI Mode select which sources to cite?

AI Mode uses a query fan-out technique that divides the original query into multiple sub-queries and retrieves sources for each independently. Only 14% of cited URLs overlap with the top 10 organic search results, confirming AI Mode uses fundamentally different selection criteria from traditional search rankings. Citation results are also highly volatile, with only 9.2% URL consistency across three repeated tests of the same query.

How many links does a typical AI Mode response contain?

SE Ranking's August 2025 research found that the average AI Mode answer contains 12.6 links. AI Overviews link to an average of 13.3 sources. Both figures are significantly higher than a traditional featured snippet, which typically cites one source.

Do top-ranking pages get cited in AI Overviews?

They're more likely to be cited, but it's no longer the norm. In July 2025, 76% of AI Overview citations came from pages ranking in the organic top 10. By March 2026, that figure had fallen to 38%, meaning 62% of AI Overview citations now come from pages outside the top 10. Ranking is still a positive signal but it no longer determines citation outcomes reliably.

How does FirstMotion measure and improve AI visibility for clients?

We audit AI citation footprints across AI Mode and AI Overviews, map citation gaps against competitor performance, and identify the specific content and technical issues causing invisibility. We then build targeted GEO programmes addressing topical depth, schema markup coverage, entity clarity, and content freshness across all key pages. Our GEO work explains the full approach.

Does AI Mode personalise results for individual users?

Yes, when users opt in. AI Mode references past search history, location data, and activity from the Google app and Google Maps to personalise responses. The same query produces different cited sources for different users based on their personal context, which is why content that speaks to specific use cases and buyer stages earns more AI Mode citations than generic overview content.

‍

Before FirstMotion, I built and exited two platforms, Obby and Baluu, and earned a Forbes 30 Under 30. Those years in founder circles gave me a close-up view of how badly search and AI discovery can be handled, even by companies with genuinely strong products.

When AI started reshaping how B2B buyers build shortlists, I launched FirstMotion with Alex Price, an exited agency founder and investor. We kept seeing the same problem: strong B2B software brands being underserved by agencies that hadn't adapted. So we built ContextualJourney™, combining audience intelligence, buyer journey mapping, and prompt mining into a single platform.

What follows covers 10 agencies in detail, the criteria we used to evaluate them, and a stage-by-stage framework to help you match your brief to the right type of partner.

Top GEO agencies in the UK: quick overview

What AI search optimisation means in 2026

The terminology is genuinely confusing. GEO, AEO, AI SEO, LLMO: agencies use these interchangeably, and some use all four simultaneously. Here's a quick breakdown:

Large language models select which sources to cite based on entity clarity, content structure, and third-party authority signals. Unlike ranking web pages in traditional search, generative AI platforms assess how well a source directly answers the query.

What separates a real GEO programme from rebadged SEO

A genuine AI search programme measures citation as a primary metric, runs real prompts through ChatGPT, Perplexity, and Google Gemini, and connects results to pipeline outcomes. GEO strategy can't be measured by organic traffic or search performance in traditional search engines alone.

The commercial case

• AI referral traffic to retail sites grew 693% year on year during the 2025 holiday season (Adobe Digital Insights, January 2026)
• AI referral sessions grew 700% across the full year 2025, the fastest-growing traffic source on the web (Previsible, 400+ tracked websites)
• AI referral traffic grew 340% year on year as of January 2026, based on Conductor's analysis of 21.9 million queries

Unlike traditional SEO, GEO focuses on how pages are retrieved and synthesised by generative engines, not just indexed and ranked. Our GEO vs SEO guide covers the full distinction.

How we selected the best generative engine optimisation agencies

No agency paid to appear. Every entry was assessed against three criteria. The right GEO agency depends on fit: your sector, your stage, and whether AI search visibility needs to stand alone or sit inside a broader programme.

Named methodology and prompt-level tracking

Structured data, entity optimisation, and content architecture for AI extraction are the baseline. Prompt-level tracking and citation reporting across ChatGPT, Perplexity, and AI Mode are the differentiators. Agencies without a named methodology are rebranding existing SEO services.

Citation outcomes, not traffic

Can they show citation results for clients, not just traffic improvements? Digital PR and GEO need to work as one: agencies that treat them as separate service lines consistently deliver weaker results in both.

B2B sector understanding

Consideration-stage queries like "best [category] software for [use case]" are where AI search is reshaping B2B pipeline. Agencies without B2B experience miss the nuances of multi-stakeholder buying cycles.

The 10 best UK agencies for AI search and GEO in 2026

1. FirstMotion

Best for: B2B SaaS and software companies at Series A-B stage with long sales cycles, complex buying committees, and pipeline goals.

FirstMotion's ContextualJourney™ platform was built around a gap most software companies don't know they have: their buyers are building shortlists through ChatGPT and Perplexity before ever visiting a website, and those shortlists often don't include them.

GEO for B2B software is not a category where a standard agency model holds up. Buying cycles are long, buying committees are senior, and the way a CISO or Head of RevOps uses AI tools to evaluate vendors is specific to the category, the moment, and the competitive set. The same senior people who set the strategy are in each FirstMotion engagement week to week, which means understanding of the client's buyers, category, and competitive position builds continuously rather than being interpreted by layers of the account team.

ContextualJourney™ is how FirstMotion structures that work. The team maps where clients appear across AI search platforms, using prompt data, ICPs and sales transcripts to build a precise picture of how buyers research and shortlist. Engagements are built around that: entity and schema audits, AI search monitoring, structured content development, and digital PR for citation authority, sequenced around the actual buying cycle. Reporting ties to pipeline from day one, with one question driving everything: is AI visibility generating opportunities?

In one B2B SaaS engagement, FirstMotion delivered a 200% improvement in AI visibility and shifted 40% of inbound enquiries to organic and AI search combined.

FirstMotion also runs digital due diligence for investors and PE firms, assessing how visible portfolio targets are across generative platforms before acquisition or growth investment. No other agency on this list offers that.

FirstMotion works with a focused number of clients at any one time. It's worth confirming availability before investing time in the process.

2. Rank4AI

‍

Best for: Businesses that want AI search visibility as a standalone programme, separate from traditional SEO or paid media.

One thing and one thing only is what Rank4AI does: dedicated AI search visibility. No traditional SEO retainer, no paid media, nothing else. Every engagement starts with an audit across six AI platforms, using a 17-section assessment that covers entity signals, content architecture, ecosystem presence, and cross-platform consistency. The methodology draws on data from over 1,400 UK business audits, which gives it a practical evidence base rather than theoretical frameworks.

Three service paths are available: Ecosystem (building AI presence outside your website, from £800/month), Full Agency (includes direct site work, from £1,500/month), and Advisory for teams that want to future proof their AI search strategy without full outsourcing. Founded by Adam Parker, the approach is systematic and the pricing is unusually transparent for a specialist generative engine optimisation agency.

Rank4AI's exclusive AI search focus is its clearest strength and its natural constraint. If your brief includes integrated SEO, content production, or digital PR, you'll need additional partners.

3. Found

‍

Best for: Larger brands that need AI search visibility tracked and reported as part of a broader performance marketing programme.

Everysearch™ is Found's trademarked framework for tracking brand visibility across generative AI platforms, social search, and traditional search engines in one place. The engine behind it is Luminr, their proprietary AI-powered platform, which maps how a brand appears wherever buyers are searching. As a full-service digital marketing agency, Found's SEO, digital PR, data, and paid media teams operate as a connected system rather than separate service lines, which is where they perform best: when AI visibility needs to sit inside a broader performance marketing agency brief. Clients include Puma, Toolstation, Fender, and House of Marley.

GEO work covers entity optimisation, schema and structured data implementation, metadata strategy, and content built for AI extraction. The infrastructure Found has built is genuinely substantial, and it's better suited to brands with the scale and budget to use it fully.

Found's model is built for scale. Brands with more focused briefs or tighter budgets will get more specialist attention from smaller partners.

4. Impression

‍

Best for: B2B and SaaS brands that want GEO integrated with digital PR, technical SEO, and genuine senior engagement across the team.

B Corp certified and independently owned since its founding in 2012 by Aaron Dicks and Tom Craig, Impression operates across Nottingham and London with dedicated sector teams for B2B, SaaS, and fintech. That vertical depth shapes how GEO gets done: knowing how buyers in those sectors research and shortlist is what determines which prompts to target and which content formats earn AI citations. Their 2024 Digital Agency of the Year win at the Global Agency Awards and a 4.5-day working week both point to an agency that's thought carefully about how it operates.

GEO services are built around earning citations through authority: digital PR and brand mention outreach sit alongside entity optimisation, schema implementation, and authoritative content structured for AI extraction. The combination of strong technical SEO and earned media capability gives them a genuinely joined-up approach to the two things AI systems assess: content quality and source credibility.

Impression is multi-channel by design. If you need a GEO-only brief or a boutique engagement model, this isn't the natural fit.

5. Passion Digital

Best for: Brands wanting GEO alongside paid media, content, and cross-channel performance, particularly B2B and professional services.

Four consecutive years as a Google Premier Partner (2023 to 2026) puts Passion Digital in the top 3% of Google's agency partners globally. The 2025 acquisition by Pixis.ai, a US AI technology firm, accelerated their AI capability: they now operate as part of Stellar, an AI-native global agency network, with access to AI forecasting tools and real-time optimisation infrastructure most independent agencies can't replicate. Named clients include Nutanix, OneTrust, Octopus Investments, Knight Frank, and Moore Kingston Smith.

The GEO offering covers entity optimisation, AI Overview optimisation, LLM performance tracking via their proprietary Deep Research methodology, semantic enhancement, and cross-platform AI search monitoring. Separating those workstreams rather than bundling them makes reporting more honest and makes it easier to see what's moving across AI search platforms and traditional search.

Passion Digital's broad service range works well for brands that want everything handled in one place. For focused GEO specialist work, you may find more depth elsewhere.

6. Blue Array

‍

Best for: Established brands and scale-ups that want the depth of a specialist organic search consultancy with a growing GEO capability built on top.

Simon Schnieders built Blue Array in 2015 after leading SEO at Zoopla, MailOnline, and Yell. What he created is deliberately different from a standard SEO agency: the Consulgency® model (trademarked) blends senior consultancy strategy with agency-scale execution. Clients include RAC, Simply Business, Funding Circle, and GoCardless. Schnieders runs the LondonSEO Meetup and authored the In-House SEO book series, which Amazon lists as a bestseller. The agency is B Corp certified, has strong technical SEO expertise, and operates from Reading and London.

Generative engine optimisation services cover AI sentiment analysis, citation gap analysis, and structured reporting across major AI models. Their technical expertise in organic search strategy underpins the GEO delivery. The Ignite package for startups gives Blue Array a broader entry point than most at this level.

Blue Array's model is strongest for brands that want senior strategic direction alongside delivery. It's less suited to a narrow AI-search-only brief.

7. Tilio

Best for: Brands that already have SEO covered and need specialist AI search measurement, tracking, and practical optimisation as a distinct programme.

A UK AI search agency based in Exeter, Tilio starts where most GEO agencies finish: measurement. Work begins by building a prompt set around your services, buyers, competitors, and decision-stage searches, then tracking how your brand appears across the major AI search platforms. Profound is the primary AI visibility data source, with Peec AI, Ahrefs, and Semrush feeding into a client dashboard that shows citation signals, competitor movement, and content recommendations in one place. Pricing is published from £499/month.

The focus is understanding whether your brand is being mentioned, cited, accurately described, and fairly compared in AI-generated responses, then improving the specific signals most likely to influence each of those factors. It's a future-proof approach for brands that want AI search visibility to compound over time.

Tilio isn't a full-service agency. Content production, link building, and technical SEO at scale are outside what they're built for.

8. Varn

‍

Best for: In-house SEO teams and technically minded marketers with complex websites who need GEO built on solid information architecture.

Where most GEO agencies lead with content strategy, Varn starts with structure. A Bristol-based Google Premier Partner, the approach to generative engine optimisation (GEO) treats it as an architectural problem first: auditing how AI systems interpret a site, then rebuilding the foundations so AI crawlers can accurately parse and cite the brand. That sequencing, structural work before content, is what separates GEO that compounds from GEO that stalls.

Services cover entity modelling, schema markup, content structuring for AI clarity, digital PR for citation authority, and AI visibility tracking across AI-powered search engines and generative search environments. Varn publishes a free guide to AI visibility that reflects a transparent, education-led approach to the discipline.

Varn's strength is technical depth. Brands that also need high-volume content production alongside structural work may need a broader partner.

9. Buried Agency

Best for: Scale-ups and growth-stage brands wanting an ROI-led approach that treats GEO and traditional organic search as a single integrated programme.

Among the first UK agencies to position explicitly around generative engine optimisation as a core organic search strategy rather than an add-on, Buried is a Bristol-based agency covering GEO, SEO, digital PR, and link building under one roof. The founding conviction is that AI search visibility and traditional organic performance aren't separate problems: brands need visibility across both traditional search and ai driven search engines to future-proof their discovery. GEO services focus on entity clarity, structured data, and content architecture for AI extraction, while digital PR and link building build the third-party citation footprint that AI systems use to assess credibility.

Small by design, which means direct access to senior practitioners rather than account management layers. A free GEO audit is available before committing to a retainer.

Being a smaller agency is a genuine advantage for some clients and a real constraint for others. Capacity during busy periods is worth discussing early.

10. ClickSlice

Best for: Ecommerce and retail brands wanting a well-established London agency that has built GEO, AEO, and LLM optimisation into its core search offering.

Joshua George's ClickSlice is a london based seo agency with unusually public credentials: a UK government commission to deliver SEO training to digital teams, a Udemy SEO course with over 100,000 students, and coverage in Forbes and Entrepreneur. Search marketing services are published from £2,500/month, making ClickSlice one of the top GEO agencies at this profile level to be transparent about pricing. GEO, AEO, and LLM optimisation are offered alongside traditional SEO, combining structured data implementation, AI-aligned content workflows, and entity optimisation.

ClickSlice appears consistently in ChatGPT and Perplexity responses when buyers search for GEO agencies in the UK, which is a proof point worth noting: they've applied the discipline to themselves. Their generative engine optimisation (GEO) and AEO capability is built on top of a heritage of strong technical SEO. Their strongest documented results are in ecommerce SEO.

B2B SaaS buyers with long sales cycles and complex buying committees should ask specifically for sector-relevant case studies before committing.

Four questions to ask any GEO agency before signing

More than simply process questions, these separate agencies that genuinely work in AI search from those that have added "GEO" to a service list.

1. Can you show us a brand appearing in ChatGPT or Perplexity for a query they don't rank for on Google?

This is the most direct test of genuine GEO capability. Organic rankings and AI citations use different signals. An agency with real GEO expertise should be able to show a client appearing in AI-generated answers for a prompt where their Google rankings wouldn't explain the citation. If they can't, the programme is likely traditional SEO with updated language.

2. How do you measure share of voice in AI answers, and which tools do you use?

The honest answer involves named tools. Peec.ai and Profound are the primary platforms in 2026 for tracking how often a brand appears in AI-generated responses across a defined prompt set. Vague references to "monitoring AI search" without specifying how are a red flag. AI search visibility is now a distinct reporting category from Google Search Console data and needs to be treated as such.

3. What's your approach to building citation authority through third-party sources?

Authority signals significantly impact AI citation selection. Brands appearing consistently in authoritative third-party publications, directories, and review platforms earn far more AI citations than brands optimising only their own content. Ask whether digital PR and citation building is part of the programme or sold separately, and ask to see examples of the third-party placements they've secured for clients.

4. Have you worked with companies in our specific vertical, and what did success look like?

GEO for a B2B SaaS company with a nine-month sales cycle is different from GEO for an ecommerce brand. The prompts buyers use, the buying committee structure, and the AI platforms they rely on all vary. Generic case studies showing traffic improvements without connecting to pipeline or revenue aren't sufficient evidence for a business-critical investment.

What separates GEO-native agencies from SEO shops with a new name?

There are now dozens of UK agencies offering AI search optimisation services. Most are applying traditional SEO thinking to a different surface, rebranding existing SEO services as GEO, and calling it generative engine optimisation. Three tests separate the genuine ones.

1. They report on AI citations as a primary metric

Not as a derivative of organic rankings. A genuinely GEO-native agency can tell you a brand's share of citations in ChatGPT for a specific prompt cluster, how that share has changed over 90 days, and which structural changes drove the movement. A digital marketing agency that's rebranded its existing SEO services can't.

2. They understand digital PR differently

In traditional SEO, digital PR builds backlinks that influence ranking web pages in Google. In generative search, it builds brand mentions in authoritative content that AI systems retrieve from and are trained on. The mechanism is different. GEO agencies that haven't made that distinction in their thinking haven't made it in their delivery either.

3. They can produce an AI visibility report

Not a screenshot of a ChatGPT response. A structured document showing which prompts were tested, which AI search platforms were checked, where the brand appeared and where it didn't, and what changed between reporting periods. That's the clearest evidence a GEO agency is running a genuine AI search programme across both AI-powered platforms and traditional search.

How to match your growth stage to the right agency

Company stage is the most reliable guide to which type of GEO agency will deliver best. Generative engine optimisation services vary significantly by scope, from foundational audit work through to full programmes covering content strategy, digital PR, and technical infrastructure.

Our lane is Series A to B, B2B SaaS and software, UK and European markets. If your brief falls here and pipeline depends on AI-mediated research, that's the context the ContextualJourney™ platform was built for.

For benchmarks on what good AI search visibility looks like at each stage, our AI search benchmarks for B2B SaaS sets out what to measure and what to aim for.

Ready to build your AI search strategy?

If your B2B software brand isn't showing up when buyers run shortlisting prompts in ChatGPT or Perplexity, you're losing pipeline at the earliest stage of the ai driven search research cycle, before a competitor's website has even been visited.

Every FirstMotion engagement starts with a ContextualJourney™ audit: mapping the prompts your buyers actually use across AI search engines and AI-driven search, identifying where you appear and where you don't, and building a prioritised organic search strategy to close the gap. Measurable from day one and tied to pipeline from the outset.

Request a GEO audit or strategy workshop to see exactly where your brand stands in AI search and what it'll take to move.

About the author

Tom Batting

Founder, FirstMotion

Tom Batting is the founder of FirstMotion, an AI Search consultancy helping B2B brands win visibility as discovery shifts from Google to AI. A Forbes 30 Under 30 entrepreneur and multi-exited founder, Tom specialises in GEO, AEO, and AI-driven organic growth for disruptive brands.

Connect on LinkedIn

‍

Frequently Asked Questions

‍

How AI Search Engines Rank and Retrieve Websites

AI search engines use a multi-stage retrieval ranking pipeline to find, score, and surface relevant content from billions of web pages. Understanding each stage determines the difference between content that gets cited and content that never enters the candidate set.

Key takeaways:

• 96.55% of web pages receive zero organic traffic, making retrieval eligibility the first barrier to address
• Hybrid retrieval combining keyword precision and vector recall consistently outperforms either method alone
• Rerankers assign relevance scores after initial retrieval to surface the most relevant passages for answer generation
• RAG architectures transform queries before retrieval to improve match quality across all pipeline stages

We've run retrieval audits on B2B software brands that rank on page one of Google but don't appear in a single AI-generated answer. The content is strong. The problem is structural: their pages fail retrieval eligibility before any relevance scoring even starts. We built our GEO practice around fixing exactly that, and this guide covers every stage of the pipeline we work through.

What is an AI retrieval ranking pipeline?

An AI retrieval ranking pipeline is a multi-stage process designed to find relevant information from a large corpus of documents and surface the best answers to a user query. According to IBM Research, retrieval augmented generation RAG combines a retrieval phase, where relevant documents are identified from an external knowledge base, with a generation phase, where a large language model synthesises an answer from the retrieved context.

The pipeline exists because large language models have a finite context window. They can't process every document on the internet before answering a question, so retrieval systems do the heavy lifting first, narrowing billions of potential sources down to the handful of relevant chunks that fit inside the LLM's context window and carry enough relevant context for grounded answer generation.

Ahrefs' study of 14 billion pages found that 96.55% of all indexed pages receive zero organic traffic from Google. The same dynamic applies to AI retrieval: the vast majority of published content never enters a retrieval pipeline's candidate set because it fails basic eligibility requirements before any relevance scoring begins.

The stages of an AI retrieval ranking pipeline

According to NVIDIA's RAG documentation, a retrieval augmented generation pipeline operates across two main phases: an offline ingestion phase where documents are processed and indexed, and an online query processing phase where retrieval and generation happen in response to a user query.

Each stage acts as a filter. Content that fails eligibility at stage one never reaches the reranker. Content that passes every stage but lacks clear entity anchoring may still be deprioritised at the answer generation stage.

How large language models and AI systems use the retrieval ranking pipeline

As IBM Research explains, RAG combines LLM generation with external knowledge retrieval to ground model responses in verifiable, up-to-date information rather than static training data. This architecture powers AI search engines, enterprise chatbots, and tools like Perplexity and ChatGPT's web search mode. Knowledge graphs also play a role in enterprise retrieval systems, providing structured entity relationships that help AI systems interpret query intent and connect relevant context across multiple documents.

AI systems across sectors including healthcare and finance use retrieval pipelines for improved decision-making, because retrieval grounds model outputs in external knowledge rather than probabilistic prediction. A senior data scientist building a RAG system for root cause analysis in a financial services environment relies on the retrieval step to pull retrieved evidence from multiple documents simultaneously, delivering relevant context that no single document contains on its own.

Stage one: data ingestion and the embedding model

Retrieval begins offline, before any user query is processed. Source documents are broken into smaller, manageable chunks, each encoded into a high-dimensional vector representation by an embedding model. Weaviate's hybrid search guide explains that these vector embeddings capture the semantic meaning of content by converting text into mathematical representations that position similar concepts near each other in vector space.

Chunk quality at ingestion directly determines retrieval accuracy downstream. Chunks that are too large dilute the semantic signal; chunks that are too small lose the context needed for grounded answer generation. The embedding model translates both the content and the user query into the same vector space, which is what enables semantic similarity search to match relevant documents even when exact keywords don't appear in both.

For content publishers, the ingestion stage has a direct implication: structured content with clear headings, explicit entity naming, and logical paragraph boundaries produces cleaner chunks. Unstructured content, JavaScript-rendered pages, and pages with poor TTFB that AI crawlers abandon before ingestion never reach the vector database and fail the retrieval process entirely.

Stage two: query understanding and query transformation

Query understanding is the stage where AI systems interpret user intent, not just the words a user typed. ZipTie.dev's pipeline breakdown confirms that query transformation enhances retrieval quality by modifying the original query before it enters the initial search, producing multiple queries that broaden the retrieval net and improve the probability of matching relevant documents.

Common query transformation techniques include:

• Query rewriting: rephrasing the original query to match vocabulary used in source documents
• Query fan-out: generating multiple queries from the same user query to capture different phrasings of the same intent
• Query decomposition: breaking complex queries into sub-queries, each sent to the retrieval system independently
• HyDE: generating a hypothetical answer and using its embedding for retrieval rather than the original query vector

The same document can fail retrieval for one query formulation and succeed for another. Content that explicitly addresses the entities and terminology users actually use in their prompts scores better across all query transformation variants, which is why entity clarity is a stronger retrieval signal than keyword density.

Stage three: keyword search and information retrieval

Keyword search, also called lexical retrieval or sparse retrieval, is a core component of information retrieval systems. It matches query terms against an inverted index of document terms to produce an initial set of search results. BM25's probabilistic scoring model, which emerged from information retrieval research in the 1970s and 1980s, scores documents based on term frequency, inverse document frequency, and document length normalisation to rank how relevant each document is to the exact keywords in the query.

BM25 excels at exact-match retrieval: product codes, named entities, rare technical terms, and specific jargon that must appear verbatim to be relevant. Its core limitation is vocabulary mismatch: a document about "machine learning model training" won't match a query for "how to build an AI" even if both cover the same concept. Semantic search addresses this gap directly by operating on meaning rather than exact keywords.

Google's 400 billion page index is narrowed to a small candidate set per query before any ranking begins. Traditional search and AI retrieval both use this two-stage architecture: broad candidate retrieval first, precise relevance ranking second.

Stage four: vector search and semantic search

Vector search, also called dense retrieval or semantic search, converts both the user query and source documents into numerical vector embeddings and retrieves documents based on semantic similarity rather than exact keyword match. Pinecone's search guide confirms that vector retrieval finds relevant results even when queries and documents share no exact terms, capturing the semantic meaning behind user intent.

The semantic similarity calculation measures the cosine distance between the query vector and each document vector in the database. Documents positioned close to the query in vector space are retrieved as semantically relevant even when they share no exact keywords with the original query. This is what allows AI search engines to correctly retrieve a document about "cloud infrastructure optimisation" in response to a query about "reducing server costs."

For content publishers, writing about a topic using natural language that covers the concept thoroughly produces better vector embeddings than content that optimises solely for keyword density. Deep learning models produce these embeddings, and the same model encodes both documents at ingestion and the user query at retrieval time, ensuring the semantic space is consistent across both.

Stage five: hybrid search, hybrid retrieval and Reciprocal Rank Fusion

Hybrid search combines keyword precision with vector recall by running both BM25 and vector search in parallel and merging search results into a single ranked list. Weaviate's RRF knowledge card explains that Reciprocal Rank Fusion calculates a combined score for each document by summing the reciprocal of its rank position across both result lists, without requiring incompatible raw scores to be directly compared.

RRF works because it operates on rank positions rather than raw scores, solving the problem of combining BM25's term frequency outputs with vector search's cosine similarity outputs. Digital Applied's 2026 benchmark data confirmed that basic RRF (NDCG 0.7068) outperforms both BM25 alone (0.6983) and pure vector search alone (0.6953) on the WANDS e-commerce benchmark, with well-tuned hybrid variants reaching 0.7497.

Hybrid retrieval enhances retrieval quality in enterprise environments because real-world queries mix both retrieval needs. Access control requirements in enterprise systems add another layer: the retrieval pipeline must filter results based on user permissions before surfacing retrieved evidence to the user interface, ensuring relevant context reaches only those with the correct authorisation.

Stage six: re ranking, answer generation and the context window

Initial retrieval optimises for recall: retrieving a broad set of potentially relevant documents. Re ranking optimises for precision: ordering those documents by exact relevance to the specific query before passing the most relevant chunks to the language model. ZipTie.dev's pipeline breakdown confirms that rerankers assign relevance scores after initial retrieval to prioritise the best content, directly determining which passages make it into the LLM's context window.

Cross-encoder rerankers evaluate the query and each retrieved document together as a pair, producing a precise relevance score. This is more computationally expensive than the bi-encoder approach used in initial retrieval, which is why re ranking operates on a shortlist of 50 to 100 candidates rather than the full index. The trade-off is significantly higher answer quality: rerankers surface relevant passages that first-stage retrieval ranked too low to reach the context window.

Answer generation is the final retrieval step. The top-ranked chunks are assembled as retrieved context and passed to the language model, which synthesises a response grounded in that evidence. User interactions with the generated answer, including follow-up queries, dwell time, and feedback signals, feed back into iterative improvements to the pipeline's ranking systems over time.

How to optimise content for AI retrieval ranking pipelines

Understanding the pipeline is the first step. The second is building a content operation that passes every stage. Most content optimisation advice targets the answer generation stage when the more critical barriers are earlier in the pipeline.

According to Google's structured data guide, implementing JSON-LD is the recommended approach for helping AI systems understand content types, entity relationships, and document metadata across all retrieval contexts.

Traditional search vs AI ranking systems

Traditional search and AI retrieval share architectural roots but diverge significantly in what they prioritise. Understanding the differences helps brands allocate optimisation effort across both surfaces rather than assuming one strategy covers both.

As FirstMotion's GEO analysis explains, GEO requires a fundamentally different discipline from traditional SEO, demanding structured content, entity clarity, and LLM-ready formatting rather than ranking signals and backlinks.

How to evaluate retrieval pipeline performance with a golden dataset

A golden dataset is a curated set of queries with known correct answers, used to benchmark retrieval accuracy across all pipeline stages. TruLens's RAG triad framework defines three primary evaluation metrics: context relevance, which measures whether retrieved chunks match the query; groundedness, which measures whether the generated answer is supported by the retrieved context; and answer relevance, which measures whether the answer addresses what the user actually asked.

For content publishers without access to pipeline internals, a practical evaluation approach is proxy testing:

• Query AI search engines with the exact questions your target buyers ask
• Observe which sources get cited and at which position
• Audit those sources against the optimisation criteria in each pipeline stage
• Track user interactions and web analytics for AI-referred traffic patterns
• Iterate based on citation rate changes after each content update

User interactions and behaviour patterns in web analytics also reveal which content is generating AI-referred traffic and which isn't reaching the candidate set at all.

Making AI retrieval visibility work for your brand

Getting consistently cited in AI-generated answers means building content that passes every stage of the retrieval pipeline, not just producing high-quality writing. The technical accessibility requirements, entity clarity demands, and direct-answer structure that AI retrieval rewards are different from what traditional SEO rewards, and the gap between the two explains why strong Google rankings don't automatically transfer to AI search visibility.

The brands that earn consistent AI citations combine three disciplines: technical infrastructure that makes content accessible to AI crawlers, content architecture that produces clean, well-bounded chunks at ingestion, and writing that delivers direct, verifiable answers at the re ranking stage.

The AI search revolution in B2B SaaS doesn't reward one optimised page. It rewards a content operation that treats retrieval pipeline eligibility as a standard requirement across every page it publishes.

If your content isn't reaching the AI retrieval candidate set, here's where to start

Most of the B2B software brands we audit at FirstMotion aren't failing AI retrieval because their content is poor quality. They're failing because their content was built for a different retrieval architecture. Fixing the structural issues, not rewriting the content, is usually where the fastest gains come from.

If you want to know exactly where your pages are failing the retrieval pipeline and what to fix first, talk to the FirstMotion team. We'll map your content against every pipeline stage and show you where the gaps are.

Frequently Asked Questions

What is an AI retrieval ranking pipeline?

An AI retrieval ranking pipeline is the multi-stage process AI search engines use to find, score, and surface relevant content in response to a user query. It includes data ingestion, query transformation, information retrieval via keyword and vector search, hybrid fusion, re ranking, and answer generation. Each stage filters the candidate set before the language model generates its response.

What is the difference between keyword search and semantic search in AI retrieval?

Keyword search uses BM25 for information retrieval by matching exact query terms against an inverted document index, scoring by term frequency and document length. Semantic search converts both queries and documents into vector embeddings and retrieves based on semantic similarity. Keyword search excels at exact-match queries; semantic search handles vocabulary mismatch. Hybrid search combines both for consistently better results.

What is Reciprocal Rank Fusion and why does it matter?

Reciprocal Rank Fusion is a merging algorithm that combines ranked results from keyword and vector search into a single list. It works by summing the reciprocal of each document's rank position in each result list, producing a unified score across both retrieval methods. RRF consistently outperforms either method alone because it operates on rank positions rather than incompatible raw scores.

How does the LLM's context window affect answer generation?

The LLM's context window is the maximum amount of text a language model can process in a single pass. Because it's finite, the retrieval pipeline must select only the most relevant chunks before answer generation begins. Rerankers exist specifically to make this selection as precise as possible, ensuring the model receives the most relevant retrieved evidence rather than just the most recently indexed documents.

How does structured data affect AI retrieval?

Structured data helps AI crawlers identify content types, entity relationships, and document metadata at the ingestion stage. JSON-LD schema markup improves chunk boundary recognition, entity clarity, and freshness signal detection. Pages with complete schema markup are over-represented in AI citations because they're more structurally extractable at every pipeline stage.

How does FirstMotion improve AI retrieval visibility for clients?

We audit content against every stage of the retrieval pipeline, from technical accessibility and ingestion quality through to entity clarity and re ranking signals. We've worked with disruptive B2B software brands to systematically improve their citation rates in Perplexity, ChatGPT, Google AI Overviews, and other generative AI search platforms by fixing the structural issues that prevent content from entering the retrieval candidate set.

Can content with lower domain authority appear in AI-generated answers?

Absolutely. LLM retrieval prioritises information gain over link authority, which means lower-authority domains earn AI citations when their content answers queries more directly than higher-authority competitors. At FirstMotion, we've helped newer B2B software brands achieve AI search visibility ahead of established category leaders by optimising for the retrieval pipeline rather than traditional authority signals.

‍