Does GEO actually work?

The original GEO gain measures answer visibility

GEO reported improvements in answer-visibility measures; its main setup optimized a source already supplied to the generator. That result does not establish additional sales.

Historical experimental setup and metric, not a forecast for current services.

Content rewrites do not reliably improve citation order

C-SEO Bench finds limited, model- and domain-dependent gains from tested rewrites. Moving a document within the supplied context has a larger effect in its comparison.

Context order was manipulated directly. No real-world SEO intervention or customer purchase was tested.

Search-stage gains and final citations can move differently

SAGEO Arena reindexes edited documents and evaluates retrieval, reranking and answer citations separately. Its StageAware method improves some reported rates while reducing another.

Constructed pipeline; selected initial top-ten targets; reported table values without confidence intervals.

Popular page features can inherit the website’s apparent advantage

In two query sets, the Chen–Koudas preprint finds apparent citation benefits for several common page features across websites that do not persist within websites.

Observational page comparison, English US-market questions, API configurations and an October 2026 preprint. It does not show that those features never help.

Referral performance does not establish a GEO return

The Marketing Science study reports ChatGPT conversion and revenue per session above paid social but below the other compared channels in its main analysis.

Historical, observational e-commerce data. The study does not randomize GEO changes or measure all AI-influenced purchases.

Google does not require special AI markup

Google documents existing SEO fundamentals as relevant to its AI search features and says no additional technical requirements or special schema are needed.

One platform’s documentation; eligibility is not guaranteed inclusion or commercial effectiveness.

The first Atlas reading test found a ceiling, not a formatting gain

Prose, labelled bullets and bullets with added facts each produced 24/24 correct answers. With the added facts, eight more answers reached a decision, including four correct rejections.

A constructed six-case comparison repeated in twelve sessions with one requested model configuration. No search, customers or sales measured. The Playbook preserves the protocol, outputs and evaluator correction.

What would count as working?

GEO, short for generative engine optimization, covers efforts to improve how a business or its information appears in AI-generated answers. For a company deciding where to invest, the useful starting point is the customer decision it wants to influence.

A source citation, a brand mention and a product recommendation answer different questions. They can also travel separately: another website can recommend your product, while your own guide supplies background for a competitor’s recommendation. A single visibility score can conceal that difference.

“40% more visible” leaves the business question open

Pranjal Aggarwal and colleagues’ original GEO study gave a generator five sources and rewrote one of them. Its strongest headline concerned how much, and how prominently, the answer represented that source. That is a useful experiment on answer construction. It leaves open whether an edited page gets discovered in a changing web index or earns a customer.

The later C-SEO Bench, by Haritz Puerto and colleagues, asked about citation order across different tasks and models. Most tested rewrites did not reliably improve it. This is not a clean contradiction: taking more space in an answer and being cited earlier are different outcomes. Its stronger “SEO” baseline moved documents within the supplied context; it did not demonstrate the return from an SEO campaign.

An edit has to survive the search that finds it

Sunghwan Kim and colleagues built SAGEO Arena to put the missing search stages back into the experiment. An edited page returns to an index. It must then be retrieved and pass through another ranking stage before it can contribute to the answer.

Their StageAware approach changes structural fields and body text for those different jobs. The results below show why the stages need separate checks. Better visibility at the beginning does not imply that every intermediate step improves, and the final citation rate still says nothing about purchases.

What else could explain a successful page?

A familiar brand may already have more coverage, links and relevant pages. If those pages also have FAQs, an across-site comparison can credit the FAQ with an advantage it did not create.

An October 2026 preprint by Mahe Chen and Nick Koudas illustrates the problem. Across 158,099 answers in two markets, collected through an academic monitoring platform the authors operate, several common page features appeared helpful across websites but not when comparing pages within the same website. Matching the question remained more informative. The authors also found that brand names could change rankings when supplied product facts stayed fixed.

The questions were in English and concerned the US market, and the answers came through selected API configurations. This is not a global consumer sample. The observational results do not prove that formatting is useless. The practical lesson is to demonstrate the effect of the change itself, rather than copying the features of an already successful competitor.

Would the sale have happened anyway?

Suppose more people arrive from an AI assistant after you improve a product page. They may already have intended to buy. A model update, a seasonal change or new independent coverage could also have contributed. Referral analytics describe a route; they do not establish the additional business caused by your edit.

Maximilian Kaiser and Christian Schulze studied 973 e-commerce websites. In their main February–August 2025 analysis, ChatGPT referrals had higher conversion and revenue per session than paid social, but lower than the other compared channels. The sample spans continents and is concentrated in the Americas and Europe. These are historical observations, not a verdict on today’s AI traffic or on a GEO programme.

For Company A, an AI-assisted sale through a retailer might still be valuable. The decision depends on margin after costs, repeat business and which sale it replaces. Measuring only visits to the brand’s own site would miss part of the outcome; counting every referred order as new demand would overstate it.

Invest in the failure you can locate

Google says the existing SEO foundations remain relevant to AI Overviews and AI Mode, with no special AI file or schema required. Accessible, useful pages and accurate product information are sensible foundations. Eligibility alone does not tell you which additional investment is worthwhile.

Our recommendation is to follow a small set of real customer decisions over time, recording the offer, market, language, competing alternatives and evidence needed to choose. Inspect the answers and available source records. A missing citation alone cannot tell you whether a page was never retrieved or was read and omitted.

A useful test must allow the competitor to win

Atlas has now run the supplied-information part of this comparison. The companion Playbook, “Test whether AI understands your offer”, contains the working files, recorded answers and checks. It compares prose, the same facts in labelled bullets, and a third version with two newly disclosed facts from a fictional offer record.

Each condition produced 24 of 24 correct answers in these authored cases. The prose baseline already passed every check, so the comparison could not demonstrate an accuracy improvement. With the two added facts, eight more answers reached a decision, including four that correctly rejected all offers. Counting only recommendations for your brand would miss that distinction.

This is a first reading-stage test. More difficult, unseen customer questions are needed before generalizing. Public discovery requires a separate live-service study; incremental sales need a further commercial comparison. Ask which part of the customer’s decision a proposed GEO investment will change, and what observation would show that it failed.

The original GEO gain measures answer visibility

GEO reported improvements in answer-visibility measures; its main setup optimized a source already supplied to the generator. That result does not establish additional sales.

Historical experimental setup and metric, not a forecast for current services.

Content rewrites do not reliably improve citation order

C-SEO Bench finds limited, model- and domain-dependent gains from tested rewrites. Moving a document within the supplied context has a larger effect in its comparison.

Context order was manipulated directly. No real-world SEO intervention or customer purchase was tested.

Search-stage gains and final citations can move differently

SAGEO Arena reindexes edited documents and evaluates retrieval, reranking and answer citations separately. Its StageAware method improves some reported rates while reducing another.

Constructed pipeline; selected initial top-ten targets; reported table values without confidence intervals.

Popular page features can inherit the website’s apparent advantage

In two query sets, the Chen–Koudas preprint finds apparent citation benefits for several common page features across websites that do not persist within websites.

Observational page comparison, English US-market questions, API configurations and an October 2026 preprint. It does not show that those features never help.

Referral performance does not establish a GEO return

The Marketing Science study reports ChatGPT conversion and revenue per session above paid social but below the other compared channels in its main analysis.

Historical, observational e-commerce data. The study does not randomize GEO changes or measure all AI-influenced purchases.

Google does not require special AI markup

Google documents existing SEO fundamentals as relevant to its AI search features and says no additional technical requirements or special schema are needed.

One platform’s documentation; eligibility is not guaranteed inclusion or commercial effectiveness.

What would count as working?

GEO, short for generative engine optimization, covers efforts to improve how a business or its information appears in AI-generated answers. For a company deciding where to invest, the useful starting point is the customer decision it wants to influence.

A source citation, a brand mention and a product recommendation answer different questions. They can also travel separately: another website can recommend your product, while your own guide supplies background for a competitor’s recommendation. A single visibility score can conceal that difference.

“40% more visible” leaves the business question open

Pranjal Aggarwal and colleagues’ original GEO study gave a generator five sources and rewrote one of them. Its strongest headline concerned how much, and how prominently, the answer represented that source. That is a useful experiment on answer construction. It leaves open whether an edited page gets discovered in a changing web index or earns a customer.

The later C-SEO Bench, by Haritz Puerto and colleagues, asked about citation order across different tasks and models. Most tested rewrites did not reliably improve it. This is not a clean contradiction: taking more space in an answer and being cited earlier are different outcomes. Its stronger “SEO” baseline moved documents within the supplied context; it did not demonstrate the return from an SEO campaign.

An edit has to survive the search that finds it

Sunghwan Kim and colleagues built SAGEO Arena to put the missing search stages back into the experiment. An edited page returns to an index. It must then be retrieved and pass through another ranking stage before it can contribute to the answer.

Their StageAware approach changes structural fields and body text for those different jobs. The results below show why the stages need separate checks. Better visibility at the beginning does not imply that every intermediate step improves, and the final citation rate still says nothing about purchases.

What else could explain a successful page?

A familiar brand may already have more coverage, links and relevant pages. If those pages also have FAQs, an across-site comparison can credit the FAQ with an advantage it did not create.

An October 2026 preprint by Mahe Chen and Nick Koudas illustrates the problem. Across 158,099 answers in two markets, collected through an academic monitoring platform the authors operate, several common page features appeared helpful across websites but not when comparing pages within the same website. Matching the question remained more informative. The authors also found that brand names could change rankings when supplied product facts stayed fixed.

The questions were in English and concerned the US market, and the answers came through selected API configurations. This is not a global consumer sample. The observational results do not prove that formatting is useless. The practical lesson is to demonstrate the effect of the change itself, rather than copying the features of an already successful competitor.

Would the sale have happened anyway?

Suppose more people arrive from an AI assistant after you improve a product page. They may already have intended to buy. A model update, a seasonal change or new independent coverage could also have contributed. Referral analytics describe a route; they do not establish the additional business caused by your edit.

Maximilian Kaiser and Christian Schulze studied 973 e-commerce websites. In their main February–August 2025 analysis, ChatGPT referrals had higher conversion and revenue per session than paid social, but lower than the other compared channels. The sample spans continents and is concentrated in the Americas and Europe. These are historical observations, not a verdict on today’s AI traffic or on a GEO programme.

For Company A, an AI-assisted sale through a retailer might still be valuable. The decision depends on margin after costs, repeat business and which sale it replaces. Measuring only visits to the brand’s own site would miss part of the outcome; counting every referred order as new demand would overstate it.

Invest in the failure you can locate

Google says the existing SEO foundations remain relevant to AI Overviews and AI Mode, with no special AI file or schema required. Accessible, useful pages and accurate product information are sensible foundations. Eligibility alone does not tell you which additional investment is worthwhile.

Our recommendation is to follow a small set of real customer decisions over time, recording the offer, market, language, competing alternatives and evidence needed to choose. Inspect the answers and available source records. A missing citation alone cannot tell you whether a page was never retrieved or was read and omitted.

A useful test must allow the competitor to win

Our proposed Atlas comparison separates clearer presentation from new information. It compares original prose, the same facts reorganized, and a version with previously omitted but verified facts. The test should include cases where the competitor fits better, delivery is unknown or the original already works.

Success means a better-supported customer decision. Repetition, varied product order and fresh questions help distinguish a useful change from a lucky answer. This controlled reading test can examine what happens after information reaches the model. Discoverability needs a separate live-service study; incremental sales need a further commercial comparison.

The protocol is prepared but has not been executed. Until that evidence exists, we recommend a limited investigation with clear criteria for continuing, revising or stopping. Ask which part of the customer’s decision a proposed GEO investment will change, and what observation would show that it failed.