What GEO is

The question you have already asked

A customer messages you on WhatsApp. She was about to order a sea buckthorn syrup, asked ChatGPT which one to buy, and it named three brands. She wants to know why yours was not among them. You open ChatGPT and type the same question: best sea buckthorn syrup in India. You get three brands, and two of them are not the ones she saw. You ask again in a new chat and the list shifts a third time. One answer links to a marketplace listing. One links to a health blog you have never heard of. One names brands and links to nothing.

That ten-minute experience contains the whole subject of this series. There was no results page and no position to hold; the engine wrote a paragraph and decided whom to name in it. Some brands were named without a link, and some pages were linked without the brand being named. The answer changed between identical questions. And nothing in your analytics recorded that any of it happened.

Each of those four observations has a mechanism behind it, and each changes what it is sensible to do, to measure and to pay for. This chapter names the mechanisms and fixes the vocabulary. The rest of the series works through them on one real property: Adiyogi Herbal, a D2C brand owned by Oriel, the publisher of this book, which sells a Sea Buckthorn Pulp Syrup. We chose a property of our own because its results are ours to publish, including the unflattering ones.

The scenario above is a composite, not Adiyogi data. At the time of writing, the baseline measurement on Adiyogi has not run, so this chapter reports no results for it.

What this chapter answers

• What does it actually mean to "show up" in an AI answer?
• Is GEO a new discipline, or SEO with a new name and a higher price?
• If AI answers send so few clicks, why should my business care about them?
• Why does this need attention in 2026 and not later?

Behind the scenes

1.1 Definition: being retrieved, cited, or mentioned in AI answers

GEO stands for generative engine optimisation. This book's working definition is: the work of getting a brand, or its pages, retrieved, cited, or mentioned in answers written by AI engines. The engines in scope are ChatGPT, Gemini, Google's AI Overviews and AI Mode, Perplexity, Claude and Copilot. Chapter 2 covers how they differ.

The definition has three verbs because an AI answer is not one event. It is a short pipeline, and a brand can succeed at one stage and fail at the next.

How an answer gets made. This is the simplified version; Chapter 4 takes it apart.

1. A user asks a question.
2. The engine either answers from what the model absorbed during training, or runs a live search first.
3. If it searches, it rewrites the question into several sub-queries and pulls back a broad set of candidate pages. Google documents this step under the name query fan-out.
4. It reranks those candidates in stages, from cheap matching on words and embeddings to a slower, more careful judgement by a second model.
5. It selects passages rather than whole pages. Perplexity's engineering blog says it ranks at "both document and sub-document level" (Perplexity Research, 25 September 2025, as checked 1 October 2026).
6. The model writes the answer. Some sources get a link beside or under it. Some brands get named in the text.

Your brand can drop out at step 3, 4, 5 or 6, and each exit looks the same from outside: you are not in the answer. That is why "we're not showing up in ChatGPT" is a symptom and not a diagnosis.

The seven outcome words. These definitions are used unchanged through all three books.

Why the definition does not say "ranked". In classic search the outcome is a position, and position 1 is better than position 5. An AI answer has no equivalent. Google's own tool confirms it: in the Search Console generative-AI reports launched on 3 June 2026, every link inside an AI Overview shares a single position (Google). The platform that built the surface does not report a first link and a fifth link. Prominence exists, but it is a property of one generated answer, and the next answer to the same question may arrange things differently.

The first distinction: retrieval is not reranking. Think of a job opening. Retrieval is being shortlisted for interview. Reranking is how you perform once you are in the room. Advice on interview technique is useless to a candidate who was never shortlisted.

Most published GEO evidence is interview advice. It takes a page that is already in the engine's candidate set, rewrites it, and measures whether it moves up. The founding paper of the field (Aggarwal et al., KDD 2024) works this way, and Chapter 4 explains why its famous "40% uplift" cannot be read as a promise about getting found. A critical survey of 45 studies published in July 2026 (arXiv:2607.14035) puts the gap in one sentence: "no reviewed technique shows a stable, longitudinal, cross-platform causal effect on organic discoverability or downstream behavior." It is an academic preprint, not yet peer-reviewed.

What changes in practice: when someone shows you a tactic that "improves AI visibility", your first question is which stage it was tested at. If the test started with the page already retrieved, it tells you nothing about a page the engine has never pulled.

The second distinction: a citation is not a mention. A citation is a link. A mention is a word. They come from different parts of the machine, which is why they come apart.

A citation is attached to a source the engine drew on. The engine can use a fact from your page and link to you without ever writing your brand name; that is a ghost citation. A mention is produced when the model writes your name into the sentence. It can do that from a third-party page that discussed you, in which case the link goes to them. It can also do it from training memory, with no search and no link to anyone.

The best data we have on the split comes from Semrush, which sells SEO software and an AI-visibility add-on. Across 3,981 domain appearances from 115 prompts in 14 countries, covering ChatGPT, Google AI Overviews, Gemini and Google AI Mode (June 2026, MODERATE):

• 61.7% of appearances were ghost citations: a source link, with the brand name absent from the answer text.
• Gemini mentioned brands 83.7% of the time but cited them only 21.4% of the time.
• ChatGPT was close to the inverse: 20.7% mention, 87% citation.

One caution on sourcing. The same dataset was also written up in Kevin Indig's Growth Memo. That is one study reported twice, not two studies agreeing.

What changes in practice: a report that says "you appeared in the answers" has told you nothing until it says whether you were named, linked, or both, and on which engine. For a business selling a product, the two outcomes do different jobs. A mention is what the buyer carries away from the conversation. A citation is what your analytics has some chance of seeing. A plan aimed at one will not automatically deliver the other, and on the evidence above Gemini and ChatGPT lean in opposite directions.

Fidelity matters more for some products than others. Adiyogi's product is a health product, a category search engines treat as YMYL ("your money or your life"). If an engine names the syrup and attaches a benefit the brand never claimed, that is a mention and a liability in the same sentence. Chapter 9 deals with this. For now, note that being named is not the finish line. Being named accurately is.

One word to distrust. The same July 2026 survey found nine different "visibility" metrics in circulation, used as if they were one. Whenever you read "visibility" in a report, a pitch or a dashboard, replace it with one of the seven words in the table. If you cannot tell which one applies, the number cannot be compared with any other number.

1.2 GEO vs SEO vs AEO vs LLMO: one activity, four labels

The field has not agreed on its own name. A Search Engine Land analysis of LinkedIn posts found that 59% of SEO influencers use "GEO", and that fewer than one-third of thought leaders kept their terminology consistent across a year (Search Engine Land analysis, reported by eMarketer, 2 April 2026). Neither publication sells GEO services.

The attempted distinctions are thin. GEO is usually described as structuring content so that generative engines cite or mention the brand. AEO is described as structuring content so it can be extracted as a direct answer. LLMO is pitched as a broader, more technical version. In practice the people doing the work treat GEO and AEO as the same activity, and the deliverable lists under all three labels overlap almost completely.

Why one activity has so many names. A label is a budget line. When we read about fifteen agency and vendor pages in August 2026, the label tracked the seller more than the work. Established SEO agencies mostly sold "GEO" or "AI search optimisation" with their SEO delivery intact. Newer tool vendors pushed "AEO" or "AI visibility" as a separate category, which supports a separate tool and a separate budget. B2B and SaaS-focused firms leaned towards "LLMO". This is our own reading of a small sample, not a survey. Oriel also has a stake: it publishes this book and plans to sell services in this market.

The term this book uses is GEO, because it is the most widely recognised and the one the academic literature uses. That is a choice of convenience. It does not mean GEO is a different discipline from AEO or LLMO. Inside the chapters we avoid the acronym where precision matters and use the seven outcome words instead.

Then is it all just SEO? Two working practitioners, quoted by Digiday (10 March 2026), give the most useful answer. Both sell SEO or content services. They could charge more by presenting GEO as a separate discipline, and they say it mostly is not.

• Jeremy Moser, CEO of uSERP: "80 percent of GEO is good, fundamental SEO." This is a practitioner's estimate, not a measurement. Nobody has measured the ratio.
• Lily Ray, VP of SEO Strategy at Amsive: GEO is following the hype arc of AMP and featured snippets, each sold as a new discipline and each an evolution of existing search work.
• Both concede a real novel core: the way engines break a query into sub-retrievals is a different architecture from crawl, index and rank.

Google says much the same from the platform side. Its documentation states that SEO best practices remain relevant and that "there are no additional requirements to appear in AI Overviews or AI Mode" (Google Search Central, as checked 19 August 2026). AI Overviews draw on the ordinary Googlebot index. Google is describing its own system, and it also has an interest in discouraging site owners from trying to game that system.

So the accurate picture is a large overlap and a small, real difference. Here is what sits on each side.

What is the same as SEO. For Google's AI surfaces, the crawler and the index are the ones you already know. A page that Googlebot cannot reach or does not index cannot appear in an AI Overview. Ranking well in classic search is still the strongest single lever in the sources we reviewed.

What is different.

1. Query fan-out. The engine expands one question into several synthetic sub-queries and retrieves for each. Your page competes for queries nobody typed. Search Console's generative-AI reports include no query data (Google, 3 June 2026), so you cannot see them. The effect shows up in Ahrefs' data: the share of AI Overview citations that also sat in the classic top 10 fell from about 76% in July 2025 to 37.9% in January 2026 (863,000 SERPs, 4 million AI Overview URLs; MODERATE-STRONG; Ahrefs sells SEO tools). That is a halving in six months, and it coincided with the Gemini 3 rollout.
2. Passage-level selection. Engines rank parts of pages, not only pages. This is a statement about what engines do. It is not an instruction to chop your content into "chunks"; Google's documentation contradicts that as a requirement, and Chapter 12 explains why it is on the dead list.
3. Different crawlers with different abilities. OpenAI, Anthropic and Perplexity run their own crawlers, separate from Googlebot. Vercel's server-log analysis found that GPTBot, ChatGPT-User, ClaudeBot and PerplexityBot fetch JavaScript files but never execute them (Vercel, published December 2024). It is older than twelve months, with no replication at the same scale. Two later checks point the same way. searchVIU, which sells a JavaScript-SEO tool, tested 32 crawlers across 1.3 billion requests in November 2025 and found that only Googlebot and Applebot render JavaScript (searchVIU lists Google-Extended separately, but it is a robots.txt token rather than a crawler; Gemini's fetches render through Googlebot's infrastructure). And no OpenAI, Anthropic or Perplexity crawler documentation through the second quarter of 2026 describes a rendering engine. Chapter 3 looks at all three sources in detail. Vercel sells hosting for server-rendered sites, so the finding suits it commercially. If it still holds, a page that classic search renders fully may be close to empty for three of the engines.
4. A different outcome. A mention or citation inside generated text, not a position on a list.
5. A different measurement problem. A rank tracker can check a position once and report it. An AI answer varies every time it is generated. Chapter 19 covers how much.

An illustration with the running example. Take the prompt we will use throughout Book 2: "best sea buckthorn syrup in India". An engine might fan that out into sub-queries about ingredients, price, reviews and safety. We cannot see the real sub-queries, so treat this as an illustration of the mechanism and not as a finding. The consequence is real either way: a page that ranks first for the head term can be absent from the answer because other pages won the sub-queries. That is one way a top-10 ranking and an AI citation come apart.

What changes in practice.

• Keep your SEO. It is the entry ticket, and Chapter 7 treats it that way. It also buys less than it did a year ago.
• Read every GEO proposal for what is new in it. In India, one agency's pricing guide puts GEO retainers at 1.8 to 3 times traditional Indian SEO pricing (upGrowth, 2026; upGrowth sells these services and no independent pricing data exists). If four-fifths of the work is SEO, you are entitled to ask what the multiple pays for.
• Ignore the label on the proposal. Ask for the deliverable list and the outcome each deliverable is supposed to change.

1.3 Branding channel, not traffic channel

Michael King of iPullRank makes the sharpest version of this point. SEO drives referral traffic. AI search visibility is closer to a branding and awareness channel, with little direct traffic. An agency that sells GEO on traffic ROI is selling the wrong value proposition (Digiday, 10 March 2026). King runs an agency that sells these services, and the statement makes them harder to sell, which is a reason to take it seriously.

Four mechanisms sit behind the claim. Each one is independent of the others.

1. The answer is the destination. The user asked a question and got a finished answer. Pew Research tracked 900 US adults across 68,879 Google searches. When an AI summary appeared, users clicked a traditional result 8% of the time, against 15% without one. They clicked a link inside the summary 1% of the time. They ended the session 26% of the time, against 16% (Pew, 22 July 2025, STRONG; Pew is a non-profit with no product to sell). This study is more than twelve months old and has no 2026 replication. The nearest newer evidence comes from Ahrefs, which sells SEO tools: it re-ran its Search Console click-through study on December 2025 data and found the click-through rate at position 1 was 58% lower when an AI Overview was present, against 34.5% lower in its April 2025 run (Ahrefs, 2026).

The wider trend points the same way. SparkToro, using Similarweb clickstream data, puts zero-click Google searches at 68.01% for January to April 2026. Different methods give levels anywhere from 45% to 68%, and the author warns that year-on-year comparison is imperfect. Trust the direction, not the number.

2. A mention has nothing to click. This follows from the definition. When Gemini names your brand without citing you, which the Semrush data suggests happens often, the reader has your name and no link. If she acts on it, she acts later and by another route.

3. The other route hides where she came from. HubSpot lays out the usual sequence. A buyer asks an AI for recommendations. Your brand is mentioned. Days later she searches your brand name on Google, clicks your paid brand ad, and buys. Last-click attribution credits the ad. The AI answer gets nothing. HubSpot's own summary: "Perfect AI attribution isn't possible" (HubSpot; the page shows only "Updated 12 August 2026"). HubSpot sells marketing software and sells an AI-visibility product, HubSpot AEO, so it has an interest in selling the workaround. The mechanism it describes stands regardless.

4. Even the clicks that happen get mislabelled. Mobile in-app browsers drop the referrer. Users copy a link out of a chat and paste it into a browser. Platforms add a noreferrer attribute to links. In each case the visit lands in GA4 as Direct. OpenAI does append utm_source=chatgpt.com to its referral links (OpenAI, as checked 19 August 2026), which helps for one platform. Anthropic offers no publisher analytics at all. The mechanism is well understood. Nobody knows what share of AI-originated visits is lost this way, and any figure quoted for it is an estimate. The working rule in this book: a GA4 count of AI referrals is a floor, never a count.

Add one fact about volume. Chartbeat, which sells analytics to publishers, found ChatGPT referrals rose more than 200% between December 2024 and December 2025 and were still under 1% of publisher pageviews (reported by Axios, 17 March 2026). That data is from news publishers, not stores, so read it as an order of magnitude.

So what kind of channel is it? Think of the person behind the counter at a medical store. A customer asks what to take, and he names a brand. That recommendation is worth a great deal to the brand he names. It also leaves no record. The brand cannot count how often he says it, cannot see which sales it caused, and cannot buy the recommendation outright. What the brand can do is make sure he knows the product exists, knows accurately what it is, and hears about it from people he trusts.

An AI answer is that recommendation at scale, with the same properties: valuable, hard to count, impossible to attribute cleanly. Marketers already have a name for channels like this. They call it brand.

What changes in practice.

• The target. Set targets in mention rate and citation rate on a fixed set of prompts, per engine, with the number of runs stated. Book 2 builds this. Do not set a sessions target.
• The proxies. Branded-search volume and "how did you hear about us" answers are reasonable supporting signals. They are proxies. HubSpot's framework ends in an "assist credit" of 25%, and that number is a judgement call, not something derived from data. Any model of this kind should print its assumption on the first page.
• The contract. A guarantee of traffic or revenue from GEO is a promise nobody can verify, including the person making it. Treat it as a reason to walk away.
• The owner. Inside a company, this work sits more naturally with whoever owns brand and PR than with whoever owns performance marketing. That is our inference from the mechanisms above; no study in the sources we reviewed tests it.

"Not a traffic channel" is a statement about volume and about what can be measured. It says nothing about the quality of the visits that do arrive. That is the subject of the next section, and the evidence there changed direction in 2026.

1.4 Why now: the 2026 numbers and Adobe's conversion reversal

Four things changed in the eighteen months before this book was written. All four are dated, because in this field a number without a date is likely to be wrong by the time you read it.

Scale. Google announced on 11 August 2026 that Gemini had passed 1 billion monthly active users (Google; the figure is Google's own). ChatGPT was at roughly 900 million weekly active users in February 2026 (OpenAI, relayed by the financial press). Those two figures use different units, monthly and weekly, so they cannot be ranked against each other. Google's AI Overviews now appear on about 43% of searches, up from 15% a year earlier (Similarweb, "2026 Generative AI Landscape" report, May 2026 against roughly May 2025; geography and device scope are not stated in the report; Similarweb sells traffic data). For Microsoft Copilot no credible consumer usage figure exists. The sources we reviewed include almost no India-specific usage figures: the one we have is that Perplexity's monthly active users in India rose about 640% year on year in the second quarter of 2025, the quarter before Airtel announced a free Perplexity Pro bundle for its subscribers on 17 July 2025 (TechCrunch, reporting Sensor Tower data, 17 July 2025; see Chapter 2).

The squeeze on clicks from classic search. The zero-click and Pew figures in section 1.3 are one half. The other is referral volume: Chartbeat measured Google Search referrals to publishers down 34% from December 2024 to December 2025 (via Axios, 17 March 2026). Google's position, stated by Liz Reid, is that click volume is "relatively stable" (Google blog post, 6 August 2025; restated in a Wall Street Journal interview in October 2025 and on Bloomberg's Odd Lots podcast on 23 April 2026). Google published no supporting data on any of those occasions. Treat it as a position and not as evidence. Chapter 2 looks at who is losing most.

Speed. The Ahrefs figure from section 1.2 deserves a second look for its pace. The overlap between classic top-10 results and AI Overview citations halved in six months. Any tactic, benchmark or case study from 2025 describes a system that has since changed. This is the reason every number in this book carries a date, and the reason the book's own rule is to re-check anything older than twelve months. Two of the sources we reviewed, Pew and Vercel, already fail that rule and are flagged where they appear.

The Adobe reversal. This is the most important number in the chapter, so here is all of it.

Adobe Analytics tracks more than a trillion visits to US retail sites, covering over 200 of the top 2,000 US retailers. For May 2026 it reported (via Digital Commerce 360, 17 June 2026; STRONG):

• Traffic to retail sites referred by AI engines grew 138% year on year.
• Those visitors converted 54% better than visitors from non-AI sources.
• They spent 53% more time on site and viewed 23% more pages.

Twelve months earlier, the same panel showed the opposite. Non-AI traffic converted 128% better than AI-referred traffic. In one year, AI-referred visitors went from converting far worse than everyone else to converting better.

Adobe sells analytics and commerce software. It does not sell GEO services, and a finding that AI traffic matters is only loosely in its interest. The panel is large and the data is first-party, which is why this is rated STRONG.

You may have seen a bigger number attached to this study: "+1,324% since October 2024". We could not trace it to Adobe's own document. It appears only in secondary press. It is not used in this book, and it is a fair example of how a real study acquires a louder statistic on its way around the internet.

What the Adobe finding supports. One argument, and it is a good one: the visits that arrive from AI engines are few, and they are now worth more each than your average visit. For a store, that is a reason to care about the channel that does not depend on volume.

What it does not support. Five limits, each of which a careful reader should be able to state.

1. It is US retail at scale. The panel is large American retailers. Nothing in it tells us whether a small Indian D2C store sees the same pattern. Adiyogi is exactly that kind of store, which is one reason it is the running example. We have no result yet.
2. It measures referred traffic only. That is the visible minority from section 1.3: clicks that arrived with a referrer. It says nothing about buyers who saw a mention and came by another route.
3. It describes the traffic, not the tactics. The finding is that AI-referred visitors convert well. It is silent on whether anything a site does increases their number.
4. It moved once and can move again. A metric that inverted in twelve months is not a constant.
5. It is a ratio. "54% better" says nothing about how many such visitors there are. The sources we reviewed give no share-of-traffic figure for retail, so this book does not print one.

On the cause of the flip, Adobe's data shows that it happened and not why. One plausible reading is that early AI referrals were curiosity clicks, while 2026 buyers arrive after the engine has already done their comparison shopping. That is a hypothesis. Nothing in the sources we reviewed tests it.

Putting 1.3 and 1.4 together. The two sections can look contradictory. They are not. The channel sends little traffic, most of its effect is invisible to analytics, and the small visible stream converts well. The defensible case for a business is therefore this: a growing number of buyers ask an AI engine before they buy; being named accurately in that answer has value you mostly cannot measure; the part you can measure has turned from poor to good within a year; and your current analytics undercounts it.

What "why now" does not mean. It does not mean sign a retainer now. Gartner's 2026 CMO Spend Survey, covering 401 marketing leaders between January and March 2026, found 15.3% of marketing budgets going to AI initiatives in general. It did not break out GEO or AEO at all (Gartner press release, 11 May 2026; Gartner sells research and advisory services). The category is not yet large or distinct enough for the biggest budget survey to count it separately. Meanwhile the asking price in India is already ₹75,000 to ₹3,50,000 a month (upGrowth, 2026, as in section 1.2), and the July 2026 survey quoted in section 1.1 found no technique with a stable causal effect.

What the evidence does support doing now is cheaper than a retainer. Check that the engines can reach and read your site; those checks are mechanical and mostly pass or fail, and Level 2 of this book covers them. Record a baseline before you change anything, because without a before there is no after. And stop reading GA4's AI referral number as a count.

Reading a vendor report

Level 1 chapters have no implementation steps. Instead, each one ends by applying its ideas to the documents you are most likely to be handed: an agency report, a tool dashboard, a proposal.

Here is a line of the kind these documents contain. We wrote it ourselves as an illustration. The figures in it are invented and it does not come from any real vendor.

"AI visibility rose from 18% to 41% this quarter, driving a 3x increase in AI traffic. Our GEO programme delivers up to 40% uplift."

Six questions from this chapter take it apart.

1. Visibility of what? The word covers at least seven outcomes. Ask whether 41% is a mention rate, a citation rate, or a mix. Ask for it per engine, because Gemini and ChatGPT behave in opposite ways and an average of the two describes neither.

2. Measured how many times? A July 2026 study (arXiv:2607.13304, STRONG) analysed 12,933 responses across 20 brands, 8 languages and 3 models. The identity of the brand explained 1.5% of the variation in the answers. Simply asking the same prompt again explained 34.8%, and the language of the query 26.5%. When re-asking moves the answer more than twenty times as much as the brand does, the difference between two single-run scores cannot be read as a change in the brand's standing. Ask for the prompt list, the number of runs per prompt, and the raw responses. Chapter 19 explains the study, and Chapter 13 builds a prompt panel that respects it.

3. Traffic from what base, counted where? "3x" of a very small number is a small number. If the count comes from GA4 it is a floor, and a rising floor can mean more visits or better labelling. More fundamentally, a report that leads with traffic is measuring a branding channel with a traffic ruler.

4. Where does "40%" come from? Almost always from the KDD 2024 paper, quoted second- or third-hand. That paper measured reranking among documents already retrieved into a fixed context. It did not measure getting retrieved. Chapter 4 goes through its design. For now the test is simple: ask the vendor which study the figure is from and which of the seven outcomes it measured.

5. What was delivered? Run the S, N, D exercise from section 1.2 on the deliverable list. Schema markup billed as a GEO lever, an llms.txt file and "content chunking" are dead-list items. Chapter 12 has the evidence on each.

6. How old is the evidence? Adobe's conversion finding reversed in twelve months. Ahrefs' overlap figure halved in six. A deck built on 2025 statistics may be arguing the opposite of what 2026 data shows. Ask for the date of every number.

What a good report looks like. It names the outcome being measured. It publishes its prompt list and states the run count. It reports mention rate and citation rate separately, per engine, with a range and not a single figure. It keeps the raw responses and will show them to you. It makes no traffic or revenue guarantee. Where it models anything, it states the assumption. None of that is exotic. In our August 2026 review of the tool market we found no vendor publishing its prompt sets or run counts (Oriel's own review).

Key takeaways

• GEO is the work of getting a brand retrieved, cited or mentioned in AI answers, and those are three separate outcomes.
• Most published GEO evidence tests the reranking of pages an engine has already retrieved, which says nothing about getting found in the first place.
• GEO, AEO and LLMO are labels for one activity, and most of the work under any of them is SEO.
• AI answers are a branding channel: they send little traffic, and most of their effect never appears in analytics.
• In Adobe's US retail data, AI-referred visitors went from converting far worse than other visitors to converting better in the twelve months to May 2026.

Checklist

Pass or fail. Each item takes a few minutes.

• I can say which of the seven outcomes I am trying to change, without using the word "visibility".
• I have asked one engine my main buying question at least three times in new chats and saved the raw answers.
• In those answers I have marked mentions and citations separately.
• I know what GA4 shows for ChatGPT, Perplexity, Gemini and Copilot as session sources, and I have labelled the figure a floor.
• No target, contract or internal slide promises a traffic or revenue number from GEO.
• Every AI-search statistic in my current deck or my vendor's proposal has a named source and a date.
• Any statistic older than twelve months is marked for re-checking.
• I have sorted my current or proposed GEO deliverables into S, N and D, and nothing marked D is being paid for.
• I can explain to a colleague why "40% uplift" describes reranking and not retrieval.

Evidence box

Every load-bearing claim in this chapter, in order of first appearance. Ratings follow the system set out in Chapter 4.

What we don't know

• How often each engine searches. Whether a buying question is answered from training memory or from a live search decides whether a citation is even possible. The sources we reviewed give no per-engine figure for this.
• Whether Adobe's finding travels. The panel is large US retailers. There is no comparable data for Indian D2C stores, and none for small stores anywhere.
• Why conversion flipped. Adobe reports the reversal, not its cause.
• What a mention is worth. No study in the sources we reviewed links mention rate to revenue. The branding-channel argument rests on mechanism and practitioner judgement.
• Whether tools see what users see. Measurement tools query APIs. Real users are logged in, with memory, location and history. No rigorous public comparison of the two exists.
• Whether the two oldest sources still hold. Pew (July 2025) and Vercel (December 2024) have not been re-run at the same scale. Later checks of a different design point the same way.
• The ratio of SEO to new work. "80 percent" is one practitioner's estimate. Nobody has measured it.
• Adiyogi's baseline. Not yet run. Everything this chapter says about the running example is mechanism, not result.