GEO versus SEO · Topic

What GEO changes: three real differences, not three slogans.

The unit of work becomes a passage. The surface becomes one answer instead of ten links. And the payoff becomes a mention rather than a click: Pew tracked 68,879 real searches and found people clicking a source inside an AI summary about 1% of the time.

Topic in GEO versus SEO12 min readUpdated 30 Aug 20268 sources, 7 primary and 1 vendor-described

Back to the roadmap Next: What to change on Monday

On this page

The short answer

Three things genuinely change when an answer engine sits between you and the reader: the unit of competition becomes a passage rather than a document, the surface becomes one synthesised answer rather than ten ranked links, and the payoff becomes a mention rather than a visit. Two further changes follow, because results are unstable and you no longer control how you are described. Pew Research Center tracked over 900 US adults through 68,879 Google searches in March 2025 and found users clicked a result 8% of the time when an AI summary appeared against 15% when it did not, and clicked a source inside the summary about 1% of the time.¹

Key takeaways

Edit the passage, not the page. The engine quotes a span and discards the rest, so each section has to survive being lifted out with no surrounding context.
Count a citation as a brand impression rather than as traffic. On Pew's panel, about 1% of visits produced a click on a source inside the AI summary.¹
Target sub-questions, not head terms. Google says both AI Overviews and AI Mode may issue multiple related searches you never see, so one question becomes several separate retrievals.⁸
Sample repeatedly and report per engine. Repeat runs of one query overlap at Jaccard 0.34 to 0.42, and three surfaces of one company at 0.11 to 0.18.²
Assume you will be paraphrased wrongly. More than 60% of 1,600 tested responses across eight engines were incorrect, so bake each qualifier into the sentence it qualifies.⁵

The shape of it

What are the three genuine changes?

Unit, surface and payoff change the work itself. Stability and representation change the reporting and the writing.

	Classic search	AI answer engines
Unit of competition	The page	A passage; median highlighted passage ~117 words
What the user sees	Ten ranked links	One synthesised answer
Query you compete on	The one typed	Several fan-out sub-queries you never see
Payoff	A click	A mention; ~1% click a source inside an AI summary
Result stability	Stable enough to rank-track daily	Repeat runs overlap at Jaccard 0.34–0.42
Consistency across surfaces	One index, one result set	URL Jaccard 0.11–0.18 across three surfaces of one company
How you are described	Your own title and snippet	Paraphrase; >60% of tested responses contained errors

Only the first four rows change what you do on a Monday: the unit changes what you edit, the surface changes what winning looks like, fan-out changes what you target, and the payoff changes what you promise your board. The last three change how you report and how you write.

What does not appear in this table is any change to the upstream half of the work. Relevance, indexation and topical coverage are unchanged and still decide whether you are eligible at all; that side belongs to the GEO versus SEO stage and the carry-over topic. Everything below is downstream.

Change one

Why is the unit a passage instead of a page?

The unit is a passage because the engine quotes a span of text and discards the rest of the page, and on at least one surface you can see exactly which span. The largest passage-level dataset published is a year-long vendor study of 15,699,298 Google AI Mode citations across 148 industries, resolving to 4.6 million unique highlighted passages across 2.7 million pages, published in July 2026. Vendor-published; described rather than linked, per this curriculum's sourcing rule.⁷

Its numbers describe what a quoted passage looks like. 47.7% of citations were scroll-to-text highlights rather than plain links, meaning the engine pointed at a specific paragraph. The median highlighted passage was 117 words. 80% put the answer in the first sentence, about 85% were fully self-contained, meaning understandable without the surrounding page, and 48% of repeatedly-cited passages opened with an explicit question against 22% of one-off passages. The same dataset also shows ranking still governing the whole thing: pages supplying 21 or more highlighted passages had a median organic rank of 1, and among passages reused 100 or more times, 76% came from pages ranking first.

The editing consequence does not depend on trusting those percentages. Write each section so it survives being lifted out with no surrounding context, because that is the only form in which it will be used: a question-shaped heading, the answer in the first sentence beneath it, every qualifier baked into the sentence rather than the paragraph above. Openers like “this means” are structurally unquotable.

Change two

Which query are you actually competing on?

Not the one the reader typed, because the engine issues several searches of its own and you never see them. Google documents this on its own surfaces: both AI Overviews and AI Mode “may use a ‘query fan-out’ technique”, which it describes as “issuing multiple related searches across subtopics and data sources” to develop a response.⁸ One question therefore becomes several separate retrievals, and your page is judged in each of them independently.

What the reader then sees is one synthesised answer rather than ten ranked links, which retires the positional vocabulary the industry reports in: there is no position three to occupy, and being the only cited source and the seventh cited source look identical in your analytics. The eligibility rule underneath has not changed at all: Google's stated requirement for a supporting link is that a page “must be indexed and eligible to be shown in Google Search with a snippet, fulfilling the Search technical requirements”, and it says there are no additional technical requirements beyond that.⁸

Two things follow for planning. Keyword-to-page mapping stops being one to one, so build coverage around the sub-questions a topic implies rather than around the head term, because the sub-questions are what actually gets searched. And stop reading an absent citation as a ranking loss: you cannot see which sub-queries were issued, and an answer may have been assembled from queries your page was never a candidate for. Grossman et al. found an AI Overview sitting above the organic results on 51.5% of 11,500 real user queries, so this is the ordinary path rather than an edge case.³

Change three

How much traffic does an AI citation actually produce?

Far less than a ranking does, on the only independent behavioural measurement anyone has published.

8%clicked with a summary15% without one

1%clicked a cited sourceinside the AI summary

26%ended the session16% without a summary

68,879searches tracked900+ US adults, Mar 2025

An AI citation produces a brand impression far more reliably than it produces a session. Pew Research Center tracked the actual browsing of over 900 US adults through 68,879 Google searches in March 2025, of which 12,593 returned an AI summary. When a summary appeared, users clicked any result 8% of the time, against 15% when none appeared. They clicked a source cited inside the summary itself about 1% of the time. And they were more likely to stop browsing entirely: 26% of visits to a page with a summary ended the session, against 16% without.¹

That is behavioural panel data from a research institution with no product to sell, which makes it the strongest evidence in this stage. The academic literature agrees about the direction of the uncertainty: the 2026 critical survey grades the claim that citation scores predict clicks, conversions or revenue at very low confidence, its lowest grade.²

The reporting change follows directly. Count citations as impressions in a high-intent moment, next to your other brand-awareness measures, and stop putting them in the same table as organic sessions. A forecast that converts citations into visits at any assumed rate is a forecast built on the one thing the field has explicitly graded as unevidenced.

Change four

What changes in measurement because answers are unstable?

Everything you used to do with a rank tracker has to be rebuilt as repeated sampling, because the same prompt asked twice returns overlapping but different sources. The 2026 critical survey reports repeat runs of the same query on commercial engines overlapping at Jaccard 0.34–0.42, with 9–28% of repeated decisions changing on temperature-controllable surfaces, and states the consequence plainly: “visibility is a distribution… a point estimate is not a stable indicator.”²

Instability across surfaces is worse than instability across runs. Grossman et al. (SIGIR 2026) measured URL-level Jaccard similarity of 0.11–0.18 between organic Google, AI Overviews and Gemini across 11,500 queries: three surfaces of one company that largely disagree about which URLs answer a question.³ Li and Sinnamon (2024) audited 1,008 generative-search responses and found that of 355 unique domains cited across two engines, only 26% appeared in both.⁴ The survey's conclusion is that these results “refute the notion of a global GEO ranking”.

Three habits follow. Sample each prompt several times over a period rather than once. Report per engine and never blend engines into one score. And treat any difference smaller than your own measured run-to-run variance as no difference at all, because at Jaccard 0.34–0.42 most week-on-week movement is noise.

Change five

What changes because you no longer control how you are described?

You are paraphrased by a system with a measured error rate, so accuracy stops being something you own and becomes something you can only make easier. Tow Center for Digital Journalism at Columbia tested 1,600 queries across eight engines in March 2025, drawing ten articles each from twenty publishers, and found more than 60% returned incorrect answers. The spread was wide: one engine was wrong 94% of the time, producing 154 citations that resolved to error pages across its 200 prompts, while the best performer was wrong 37% of the time. Licensing agreements with the model providers gave no guarantee of accurate citation, and publishers who had blocked the crawlers were still cited.⁵

An earlier Tow Center study in November 2024 took 200 direct quotes from 20 publishers and asked one engine to identify their source: 153 of the 200 responses were partially or entirely incorrect, and the system signalled uncertainty only 7 times.⁶ Academic work finds the same weakness inside answers: Liu et al. (2023) found only 51.5% of sentences in generative-search answers fully supported by their citations, and a 2026 study classified about 11% of 98,020 atomic claims as insufficiently supported.²

You cannot fix that from your side. What you can do is reduce your exposure to it, and the technique is the same one the passage data points at: write sentences that stay true when quoted alone. Put the qualifier inside the sentence rather than in the one before it, date every claim in the claim, and name the source in the text rather than only in a footnote. A sentence that is only true in context will eventually be quoted out of it.

The honest limit of this page

The Pew figures describe Google only, US adults only, in March 2025, and the 1% is specifically the rate at which people clicked a source cited inside a summary, not a total AI-referral rate for a site, which no independent study has measured. There is no comparable behavioural panel for ChatGPT, Perplexity or Copilot, so the payoff row of this page rests on one engine and one month. The passage-level percentages are vendor-produced, correlational, and lack a base rate: the study does not report what share of all web paragraphs are answer-first, so it corroborates a plausible mechanism rather than proving one.

Where a product fits, and where it does not

Four of the five changes here are fixed by editing, not by buying: rewrite your highest-intent sections to be self-contained, build coverage around sub-questions instead of head terms, move citations out of the traffic forecast and into brand reporting, and bake every qualifier into the sentence it qualifies. Only the stability change needs a machine, because sampling one prompt across five engines several times a week by hand is a job nobody keeps doing after month two. Bavior runs a fixed prompt set on a schedule and records which sources each answer cited, so a real move separates itself from noise. It cannot correct a wrong statement about you, and it promises nothing about citations or traffic. Try the free visibility check first; paid plans are from $99/mo billed monthly, or $79.17/mo billed annually (as of 29 Aug 2026).

Sources, all checked 30 Aug 2026

Pew Research Center, “Google users are less likely to click on links when an AI summary appears in the results”, 22 Jul 2025; 900+ tracked US adults, 68,879 searches in Mar 2025, 12,593 with an AI summary: pewresearch.org
“Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023–2026)”, 15 Jul 2026, arXiv:2607.14035 (preprint); confidence table, run-to-run variance, and the Liu et al. and Xu et al. support figures: arxiv.org/abs/2607.14035
Grossman et al., SIGIR 2026; 11,500 queries; AI Overviews above the organic results on 51.5% of them; URL-level Jaccard 0.11–0.18 across organic Google, AI Overviews and Gemini: arxiv.org/abs/2604.27790
Li & Sinnamon, 2024; audit of 1,008 generative-search responses; 355 unique domains, 26% cited by both systems. Reported via the critical survey at note 2; no open-access link at the time of checking.
Jaźwińska & Chandrasekar, “AI Search Has a Citation Problem”, Tow Center for Digital Journalism, Columbia, 6 Mar 2025; 1,600 queries, 20 publishers, 8 engines: cjr.org
Tow Center for Digital Journalism, “How ChatGPT Search (Mis)represents Publisher Content”, 27 Nov 2024; 200 quotes from 20 publishers: cjr.org
Passage-level figures: a vendor study of 15,699,298 Google AI Mode citations across 148 industries over a year, resolving to 4.6M unique highlighted passages across 2.7M pages, published 29 Jul 2026. Vendor-published; described rather than linked, per this curriculum's sourcing rule.
Google Search Central, “AI features and your website”, last updated 10 Dec 2025 (first-party; query fan-out and the indexation requirement): developers.google.com/search/docs/appearance/ai-features

Continue reading

Learn GEO GEO roadmap Learn GEO Previous: Topical coverage Learn GEO Next: What to change on Monday

FAQ

Frequently asked questions.

Do AI citations bring traffic?

Far less than a ranking does, on the only independent behavioural data published. Pew Research Center tracked over 900 US adults through 68,879 Google searches in March 2025 and found that when an AI summary appeared, users clicked any result 8% of the time versus 15% without one, and clicked a source cited inside the summary about 1% of the time. They also ended the session more often, 26% of pages with a summary against 16% without. Treat a citation as a brand impression in a high-intent moment rather than a traffic source.

What is the unit of optimisation in AI search?

A passage, not a page. The largest passage-level dataset published is a year-long vendor study of 15.7 million Google AI Mode citations, resolving to 4.6 million unique highlighted passages across 2.7 million pages. It found 47.7% of citations were scroll-to-text highlights pointing at a specific paragraph, with a median highlighted passage of 117 words, 80% putting the answer in the first sentence and about 85% fully self-contained. The practical rule that follows does not depend on those percentages: write each section so it survives being lifted out with no surrounding context.

Which query does an AI answer engine actually search for?

Several of its own, not the one that was typed. Google documents that both AI Overviews and AI Mode may use a "query fan-out" technique, which it describes as issuing multiple related searches across subtopics and data sources to develop a response. One question therefore becomes several separate retrievals, and your page competes in each independently. Two consequences: build coverage around the sub-questions a topic implies rather than around the head term, and stop reading an absent citation as a ranking loss, because you cannot see which sub-queries were issued.

Why do AI visibility numbers move so much week to week?

Because the measurement itself is noisy, not because your visibility changed. The 2026 critical survey reports repeat runs of the same query on commercial engines overlapping at Jaccard 0.34–0.42, with 9–28% of repeated decisions changing, and concludes that "visibility is a distribution… a point estimate is not a stable indicator." Across surfaces it is worse: URL-level Jaccard of 0.11–0.18 between organic Google, AI Overviews and Gemini across 11,500 queries. Sample each prompt several times, report per engine, and treat any move smaller than your measured variance as nothing.

Can I control how an AI engine describes my brand?

No, and the error rates are high enough to plan around. Tow Center at Columbia tested 1,600 queries across eight engines in March 2025 and found more than 60% returned incorrect answers, with one engine wrong 94% of the time, producing 154 citations that resolved to error pages across its 200 prompts. Licensing deals with the model providers gave no guarantee of accurate citation. What you can control is your exposure: bake every qualifier into the sentence it qualifies, date claims inside the claim, and name sources in the text, so a passage quoted alone is still true.

What should I stop reporting now that answers replaced links?

Three things. Stop forecasting sessions from citations, because the 2026 critical survey grades the claim that citation scores predict clicks, conversions or revenue at very low confidence, its lowest grade. Stop reporting one blended AI visibility score, because URL overlap between three surfaces of a single company runs at Jaccard 0.11–0.18. And stop treating a single prompt run as a measurement, because repeat runs of the same query overlap at Jaccard 0.34–0.42, which makes most week-on-week movement indistinguishable from noise.

Bavior Editorial

The team that researches and maintains Bavior’s writing on Reddit marketing and AI search visibility. Every figure here is attributed to a named source with the date it was checked, and none of our links are affiliate links.

Found a number that looks wrong? Tell us and we will re-check it: support@bavior.com