AI search statistics 2026: research and limitations

Imagine asking ChatGPT, 'Which accounting software is best for a small business?' You can compare options, learn their differences and choose a shortlist before opening a company's website. AI search can shape that decision even when it sends no immediate visit.

The statistics in this guide explain different parts of that journey. Adoption figures show how many people use AI products. Click and referral studies show when they visit a website. Studies of the answers show which brands get mentioned, recommended or supported by a source.

Parse, the publisher of this guide, studies how brands appear in AI answers. We brought our research together with studies and announcements from Google, OpenAI, Pew Research Center, Adobe and others. Each number below keeps its study dates and the group it describes, so you can judge where it applies.

AI search means using an AI system to find, explain or compare information. A prompt is the question you type. The answer may name products, recommend an option and include links to the pages used as sources. Those links are called citations.

Generative engine optimization (GEO) means improving how your content or brand appears in AI-generated answers. Answer engine optimization (AEO) covers visibility in direct answers more broadly, including search features that existed before today's chatbots. The terms overlap in practice. This guide focuses on outcomes you can measure, rather than which label you use.

The products also differ. Google AI Overviews are AI summaries shown within Google's search results. AI Mode is Google's dedicated AI search experience. ChatGPT handles many tasks; ChatGPT Search specifically refers to answers that use its web-search feature. A study of one should not automatically be applied to the others.

Four outcomes matter: a mention puts your name in an answer; a recommendation presents you as an option; a citation links to a source; a referral sends someone to a website. For example, 'Tool A integrates with Tool B' mentions both products but does not necessarily recommend either. 'Choose Tool A for a small team' is a recommendation. A linked review is a citation, and a click that brings someone to your site is a referral.

Before comparing percentages, ask 'a percentage of what?' Twenty out of 100 answers mentioning a brand is different from 20 out of 100 source links pointing to its website. The table below explains the main measures used in this guide.

OpenAI reported that ChatGPT reaches 1.2 billion people each week on October 5, 2026. Google reported more than 2.5 billion monthly active users for AI Overviews and more than 1 billion for AI Mode at its May 19, 2026 keynote. These are company-reported usage figures.

They establish the scale of the products. They do not establish a combined audience. Weekly use cannot be compared directly with monthly use, and the same person may use every product in the table. ChatGPT also handles writing, coding and other tasks; its entire user base is not an AI search audience.

For a marketing team, the practical question comes next: which of these surfaces answers the questions your customers ask? Broad adoption supports taking measurement seriously. It does not tell you how much to spend in a particular category, country or platform.

Semrush tracked more than 10 million keywords during 2025. AI Overviews appeared for 6.49% in January, 24.61% in July and 15.69% in November. Coverage grew and then fell; it did not rise steadily throughout the year.

These are shares of tracked keywords, not shares of all searches people made. In a keyword study, an obscure phrase and a popular phrase can each count once. That makes the result useful for understanding where AI Overviews appear, but different from a study of actual browsing.

For a website owner, the next question is whether AI Overviews appear for the searches that bring customers to the site. The overall rate can fall while AI answers become more common for your particular topics.

Conductor's 2026 AEO / GEO Benchmarks Report found that identifiable AI referrals accounted for 1.08% of website sessions across its industry sample. On average across those industries, ChatGPT supplied 87.4% of AI referrals. The analysis covered 1,215 enterprise customer websites and more than 3.3 billion sessions from May through September 2025.

This measure leaves out an important part of Google. Visits from AI Overviews and AI Mode were counted with ordinary Google search traffic, so Conductor could not separate them into the AI-referral total. The 1.08% figure is therefore not a measure of all visits influenced by AI.

The industry breakdown shows why one average is a poor target for every website. The recorded AI share was 2.8% in information technology and 0.35% in utilities. These are historical observations from the report's sample, not current targets for your site.

Small referral numbers do not settle how much AI affected a purchase. Someone may read an answer and visit a site later by typing its name. Another person may get everything they need without visiting at all. Website analytics show the visits you received; reading the AI answers helps explain what happened before them.

Pew Research Center studied browsing by 900 US adults in March 2025, covering 68,879 unique Google searches. People clicked a traditional search result on 8% of visits with an AI summary, compared with 15% without one. Put simply, about 8 of every 100 search visits with a summary included a click on a traditional result, versus 15 of every 100 without one.

Links inside the AI summary received a click on just 1% of visits with a summary. Browsing sessions ended after 26% of summary visits, versus 16% without one.

Pew recreated the result pages in April, after observing the browsing in March. It did not randomly assign people to see summaries, and different types of searches can lead to different click behavior. The study shows an association, not proof that the summary alone caused the entire gap. The difference between 8% and 15% is seven percentage points.

A search that produces no click is often called a zero-click search. This predates AI summaries: SparkToro and Datos reported no clicks for 58.5% of US Google searches and 59.7% of EU searches in their 2024 study. Those rates cover the sampled Google searches overall, not AI answers specifically.

A top Google ranking can receive fewer visits even when it keeps its position. Ahrefs' February 2026 study found that AI Overviews were associated with a 58% lower average click-through rate for the top-ranking page. Click-through rate, or CTR, is the share of search impressions that result in a click.

Ahrefs studied 300,000 keywords: half with AI Overviews and half informational keywords without them. It compared desktop Google Search Console data from December 2023 and December 2025. On the AI Overview keywords, the top result's CTR fell from roughly 7.3% to 1.6%. In the comparison group, it fell from 7.6% to 3.9%.

The reported 58% accounts for the decline in the comparison group. Ahrefs used that group to estimate what CTR might have been without an AI Overview, then compared the observed result with that estimate. It is not a claim that every website lost 58% of its traffic.

For your own site, check clicks alongside rankings. If a page still ranks first but receives fewer visits, look at the search results themselves: an AI answer, a different layout or a shift in the kinds of searches can change how often people click.

Adobe's Q2 2026 report found AI-referred visitors to US retailers converted 42% better than non-AI visitors in March 2026 and generated 37% more revenue per visit. These are relative comparisons in Adobe's US retail analytics, not percentage point gains or results for every industry.

A relative improvement is different from a percentage-point gain. For illustration, a 2% conversion rate that improves by 42% becomes 2.84%, not 44%. Those example rates are not Adobe's reported baseline; they show how to read the comparison.

The same report recorded first-quarter year-over-year growth in AI visit share of 393% for retail, 233% for travel, 158% for financial services, 84% for media and entertainment, and 63% for technology and software. Rapid growth from a small starting share can coexist with a modest contribution to total traffic.

One possible explanation is that people compare options inside an AI answer before clicking. The study does not prove that explanation. The practical question is whether your AI-referred visitors buy, sign up or become useful leads. Compare those outcomes with other visitors, while keeping the number of visits in view.

Ranking well in Google and being cited in an AI answer are related questions, but they are not the same result. In a study published in March 2026, Ahrefs examined 863,000 keyword result pages and 4 million AI Overview URLs. Of the cited URLs, 37.9% appeared in the first 10 search-result blocks. Counting only ordinary organic links changed the figure to 37.1%; blocks can also include features such as ads and video packs.

Parse asked a different question: could five independent web-search services find the same pages that ChatGPT Search and Google AI Mode cited? The services' combined top-10 results included 18.0% of ChatGPT Search's cited pages and 36.9% of Google AI Mode's. This comparison covered 214 buyer questions and 1,370 question-and-day observations from September 14 to October 6, 2026.

These results use different search systems and should not be combined. The Parse comparison was not a test of Google organic rankings, and the AI answers and comparison searches could be up to three days apart. Google AI Mode's recorded source lists also became much shorter after August 29; the study could not tell whether the product or the collection method caused that change.

A useful next step is to inspect the sources in the answer you want to appear in. A page may rank well for the original query while the AI cites a page that answers a more specific follow-up question.

An AI search system can write its own searches before answering your question. For example, after a broad question about accounting software, it might look for a particular brand's prices or integrations. The buyer does not have to name that brand first.

In 41,334 ChatGPT Search gpt-5-6 answers that named brands, 79.1% searched for a brand the user had not named. These answers were recorded from September 29 to October 6, 2026. Searched brands were named first 23.4% of the time, compared with 10.9% for unsearched brands, about 2.1 times as often. The full study covers 691,961 recorded queries across two periods; the 79.1% result belongs to the recent group, while January–March models showed shares near 16.7%.

The related site-search study examined 155,899 searches that ChatGPT wrote in that September–October window. Of those, 32.5% used site: to search within one website. Within answers that ran at least one such lookup, brands looked up had their own site cited 85.4% of the time. Brands named without a lookup in those same answers had their own site cited 45.7% of the time.

This association does not prove that being searched caused a citation or recommendation. ChatGPT may look up a brand it already plans to discuss. It does give a practical reason to make prices, integrations and product details easy to find on your own website. Earlier recordings had collection limitations, so this short window cannot establish when the behavior began.

Third-party coverage matters, but a brand's own site is far from irrelevant. Parse's May–July 2026 analysis found 27.2% of eligible links between ranked brands and cited pages pointed to the recommended brand's own website. That study counted 1,060,513 links and matched ownership at the main website address.

A broader publisher study covering May 24 to August 23 found vendors supplied 48.6% of recommendation-backing citations, including both the recommended brand and other sellers in the market. Commercial media supplied 20.0%. ChatGPT Search drew 50.4% of its recommendation evidence from the recommended brand's own site.

These numbers answer different ownership questions. 'Vendor' includes more than 'the recommended brand.' The broader study also left 18.1% of links unclassified by publisher type. Treat the table as separate measured findings, not parts of one pie.

The practical implication is to audit both places where a buyer could encounter evidence: your own pages and the independent pages that already appear in answers. Choosing only one leaves part of the observed citation system unexplained.

In Parse's homepage study, 67.6% of 289,494 brand-owned citation links pointed to a deeper page. A homepage was a URL with no path after the hostname. Product pages, documentation, pricing and other paths belonged to the deeper-page group.

That is a useful counterweight to homepage-only optimization. A category introduction can establish who you are, while the page that answers a particular question supplies the evidence. The full page-type studies found comparison-labeled pages in 24.4% of cited answers, review-labeled pages in 20.0%, product-labeled pages in 18.8%, pricing or cost pages in 6.4% and FAQ-labeled pages in 3.4%.

Those page-type rates come from their own studies and can overlap within an answer. The classification uses page labels rather than an experiment that changes a page's format. It does not show that adding an FAQ or changing a URL will increase citations.

Start with the information a prospective customer needs to make a decision. Give that information a clear, accessible page and keep it accurate. Then measure whether the page appears for the questions it was intended to answer.

Parse counted 7,433,783 citation links in answers to buying questions from May 24 to August 23, 2026. YouTube supplied 5.0% and Reddit 4.1%. A group of 13 other publishing platforms, including LinkedIn, Medium, Quora and Wikipedia, supplied 4.2% combined. All other websites supplied 86.7%.

Within that group, LinkedIn supplied 70,610 links, Medium 65,138, Facebook 60,675 and Instagram 48,613. Wikipedia supplied 6,017, or 0.08% of all links. These are shares of links in answers to the buying questions Parse tracked. They do not describe Wikipedia's role in every kind of informational AI answer.

The engine split was substantial: the 13-platform group's citation share was 5.4 times higher on Google AI Mode than on ChatGPT Search. A distribution plan built from a combined average can miss the source behavior of the platform you care about.

A separate study following the same questions found Reddit's historical share fell from 4.9% to 3.8% between its measured periods. A website can remain widely cited while losing share as the rest of the citation set grows. Neither link share nor answer coverage is evidence that publishing on the site causes a brand recommendation.

In Parse's directory study, G2 or Capterra was the leading source in none of 835 categories. The median category had 260 distinct source domains, and only 14.2% of brands had any directory citation in that study. Those figures argue against treating one software directory as the universal source of AI recommendations.

A later market-level comparison found G2 received 4.6 times as many citations as Capterra, while review sites together supplied 1.2% of AI citations. In 32 of 41 measured markets, one review site held at least half of review-site citations. A small overall share can conceal a concentrated role within a specific market.

Budget decisions should therefore start with the pages AI already cites for your buyer questions. A directory listing can matter where that directory is consistently part of the answer. Its overall prominence or paid marketing pitch does not establish the same value in every category.

Seeing your name in an AI answer does not always mean the answer recommends you. A company may appear because it is an integration partner, an example, an option to avoid or the product the person already uses.

Parse tracked 743,998 brand mentions. Of those, 72.7% were given a place in the answer's ranked recommendations, 12.9% ranked first and 35.7% had a cited source attached. The groups overlap: a brand can be recommended, rank first and have a citation in the same answer.

Even the first name in the text can mislead. In a separate study, 9.3% of answers named a different brand before their first recommendation. Another study found that 65.2% of integration-partner mentions did not also present that partner as an option. 'This tool works with Salesforce' mentions Salesforce, but may be recommending the other tool.

When checking visibility, keep these outcomes separate. A mention shows that your name appeared. An explicit recommendation shows that you were offered as a choice. First place shows which option the answer led with. Read the sentence around the name before deciding which outcome you have.

Across 227,755 answers ranking at least two brands, Parse found 16.7% cited every ranked brand. Another 54.0% cited some, and 29.3% cited none of the ranked brands. The study covered ChatGPT Search and Google AI Mode from May 24 through July 16, 2026.

This study counted answers recommending at least two brands and checked whether each recommended brand had a source attached. An answer could contain other links and still give no source for its recommendations. 'No cited brands' therefore does not mean 'no links anywhere in the answer.'

The first choice is not exempt. In a separate analysis of 215,349 answers with a top-ranked brand, 47.2% had no citation attached to that top brand. Google AI Mode attached a citation to its top brand in 60.4% of answers, compared with 44.8% on ChatGPT Search.

If you count only citations, you will miss some recommendations. For a reader, the same gap means an answer can make a clear choice without exposing the evidence behind that choice.

A cited brand appeared in the ranked recommendations in 99.3% of brand appearances that met the study rules in Parse's July–August study. Within answers that ranked brands, cited brands were picked first 22.0% of the time, compared with 10.4% for named but uncited brands. Citation and selection were associated, but a citation was far from a guarantee of first place.

A different study asked whether the brand with the most attached pages ranked first. In 59.0% of answers with a clear citation leader, that brand ranked lower. Tied leaders and answers without a cited ranked brand were excluded; 103,679 answers remained for that comparison.

These findings leave room for several mechanisms. A longer discussion can need more sources. A less familiar option may require explanation. The first recommendation may be stated briefly. The observational data cannot decide which mechanism drove an individual answer.

Track citations to understand where an answer gets its evidence, and recommendation position to understand which brand it chooses. More source links and a higher recommendation are separate improvements.

Parse matched 81,800 answer pairs to the same questions in the same collection rounds between May 24 and July 16, 2026. ChatGPT Search and Google AI Mode selected different top brands in 63.8% of pairs. They agreed in 36.2%. Parent companies and their product brands were grouped together to avoid counting a naming difference as disagreement.

This can coexist with the older finding that ChatGPT and Google AI Overviews shared the most frequently recommended brand in 93% of 1,655 categories. That older study aggregated recommendations across a six-month window. It compared different products, at a different level, over a different period.

A brand can be the most frequent winner across a category while individual answers often choose someone else. The 93% study compared category leaders; the 36.2% agreement result compared individual answers to the same question. The products and study periods also differed. Use each number for the question it actually answers.

Do not report success on one platform as coverage of AI search overall. Review the same buyer questions across the platforms your audience uses.

Parse's concentration study analyzed 631,087 recommendations across 9,658 buyer questions on ChatGPT and Google AI Overviews from October 2025 to April 2026. The median top-brand share was 33% across all questions. Consumer Electronics had an 81% median, while Sales and Marketing had 22%.

A broad benchmark therefore hides very different competitive conditions. In a concentrated category, one familiar brand may absorb much of the recommendation share. In a more fragmented category, several specialists can appear repeatedly without any one dominating.

The reported share is the leading brand's fraction of recommendations for a question. It is not market share in sales or the probability that a consumer will buy it. The medians also describe an earlier pair of products, so they should not be presented as today's category rankings.

Benchmark against the questions you compete on and against the brands actually appearing beside you. A universal target can be too easy in one category and unrealistic in another.

Parse compared 497,084 same-day answer pairs using different phrasings of the same buyer question with 106,823 rerun pairs using identical wording. Average brand-set overlap was 11.7% across phrasings and 38.0% across identical-wording reruns. In 45.3% of phrasing pairs, the two answers shared no brands.

Overlap here is the share of brands common to both lists out of all brands either list contains. It is not the share of buyers whose preference changed. The analysis covered 340 tracked questions and 1,677 phrasings on ChatGPT Search and Google AI Mode from July 1 through August 26, 2026. For example, if one answer names A and B and another names B and C, they share one of three distinct brands: 33.3% overlap. That is an illustration of the calculation, not another study result.

This is why a prompt set should represent the ways people ask, including their constraints and context. Repeating one neat category phrase can produce a stable-looking metric that says little about other plausible phrasings.

When evaluating a content change, preserve the prompt set. If you rewrite the questions at the same time, changes in the answers may come from the measurement rather than the content.

The wording effect is not only stylistic. In Parse's constraint study, questions with compliance requirements reduced how often the usual category leader was named first by 42% relative to comparable questions without that requirement. Among the changes in leadership that met the study's confidence rules, 32.4% went to brands never named first on a generic prompt. A restriction can create a different competitive field.

Geography also matters. Among answers to prompts naming a country other than the US, 62.3% of first-named brands were classified as local. That study used a conservative rule based on country concentration and website domains; hand checks showed it missed some local brands. The observed prompt set was heavily US-weighted.

Pricing and contract terms were the stated reason for 4.1% of recommendations tied to a specific buyer requirement: 38,214 of 920,930 recommendations. Feature requirements were the reason about 13 times as often. A need might be 'works offline' or 'fits a monthly budget.' The study counted recommendations linked to those needs, not distinct needs or the share of consumers who care about price.

The useful response is to cover the real decisions a customer faces: geography, integrations, deployment requirements and the constraints that disqualify an otherwise popular option. A generic 'best tools' page cannot answer every one of those questions.

Across 161,023 consecutive answer pairs, Parse found the top recommendation changed in 43.6% of cases. ChatGPT Search changed it in 44.7% and Google AI Mode in 42.6%. These answers used the same prompts on the same engine, so no deliberate wording change was required.

This was the pattern in the answers collected from May to July 2026. Some questions were more stable than others. The overall rate does not predict how often your particular question will change.

Citation changes do not perfectly track recommendation changes either. In another study, 21.1% of consecutive pairs repeated no cited page at all. Among those pairs that shared no sources, 47.3% still kept the same top brand. With two or more repeated pages, the share keeping the same top brand was 62.8%.

Check several answers before deciding that a brand has gained or lost visibility. A changed source list does not automatically mean a changed recommendation, and one changed recommendation does not establish a lasting trend.

A Parse study compared 423,341 ChatGPT answers before and after a major model update, using 16,161 prompts observed under both versions and a second model that did not change during the comparison. Sources per answer fell from 23 to 12, a reported 48% decline. That unchanged model moved by 4% in the other direction.

A later study following the same set of questions found ChatGPT Search's mean brand count per brand-naming answer fell from 6.4 to 5.2 across the August 2026 model transition, a 19% decline. On Google AI Mode, the median number of named brands instead rose from four in June to six in August. The platforms did not move together.

The model-update study also separated an update from the differences that arise when the same question is asked again. Across the earlier transition, 38.6% of answer pairs kept the same first-named brand, versus 48.2% on the model that had not changed. That is a 9.6 percentage point difference, not evidence that every changed answer was caused by the update.

If many answers change at once, check whether the platform changed its model or answer format. Your website is only one influence. Comparing the same questions over time, with those product changes noted, makes the result easier to interpret.

Daily answers move around, but familiar brands can retain an advantage across many questions. In a Parse study following the same question set over time, brands already observed in October 2025 received 64% of ChatGPT Search's first-named picks and 67% of Google AI Mode's in September 2026. Only 14.8% of brands first observed in the last six months reached first place even once.

'First observed' is not 'newly founded.' It means the brand first appeared in the study's recorded answers. The long study also spans a change from ChatGPT and AI Overviews to ChatGPT Search and AI Mode, with months with incomplete or unreliable data left out of trends. Those boundaries remain part of the result.

A separate study of market rankings found that completed periods at number one lasted a median of 12 observed days. In other words, half of those completed periods were shorter and half longer. Brands still leading at the end of the study were excluded, so 12 days is not the expected lifespan of every leader.

Together, the studies explain why a lost answer and a lost market are different events. Monitor individual recommendations for evidence and aggregate comparable observations for strategic decisions.

A source link lets you check a claim; it does not prove the claim is correct. Parse compared 150,811 pairs of statements and their cited pages. When the AI praised a brand, 59.5% of the checked pairs linked to page coverage also classified as positive. The rate was 68.9% for ChatGPT Search and 56.2% for Google AI Mode.

The primary mismatch was neutral coverage. A documentation page can establish that a feature exists without supporting an answer's claim that the product is excellent. Only 1.9% of checked praise pairs cited page coverage classified as negative; lack of positive support should not be confused with direct contradiction.

Software classified the tone of both the answer and the page. This checks agreement in positive, neutral or negative language; it does not establish that every factual claim is true. In a separate study, 38.5% of checked cited pages did not mention the brand attached to the citation at all. Opening the source is still necessary.

For your own brand, prioritize material claims: price, compatibility, security, availability and explicit comparisons. Verify the relevant passage and the product it describes, rather than assuming the citation marker makes the sentence correct.

Parse extracted 358,902 prices quoted for brands in AI answers. Only 7,594 could be compared with suitable saved copies of pricing pages. Within that small group, 62.8% matched a listed amount within 10%. The other 97.9% of all extracted quotes could not be checked by this method.

The hand review changes the interpretation. Of 100 sampled candidates, 79 were genuine brand price quotes. Of those, 65 were right, eight were wrong and six could not be settled: 82.3%, 10.1% and 7.6%, respectively. Some automated mismatches were correct prices that the crawled text did not contain.

It would therefore be wrong to turn the automated 37.2% non-match rate into an AI pricing error rate. It would also be wrong to generalize eight confirmed errors in a small, selected hand sample to every AI price quote.

If you publish prices, put the amount, billing period, plan name and country together so a reader can tell what the price covers. If AI quotes a different amount, inspect the original source and the version available to the search system. Missing page text can look like a wrong answer even when the quoted price is real.

Parse found 94.1% of descriptions attached to first-ranked brands were positive in its rank-and-praise study. Yet 84.7% of appearances of ranked brands that received criticism also carried praise in the mixed-language study. An answer can recommend a product and still point out an important limitation.

Across comparisons of the same ranked brands between ChatGPT Search and Google AI Mode, 91.6% of cases with criticism carried that criticism on one engine only. In the repeated-answer criticism study, 74.0% of matched cases carried criticism in only one answer. A single negative sentence is therefore weak evidence of a persistent cross-platform reputation problem.

Language is also unstable at the word level. In the repeat-description study, 78.45% of comparisons of repeated descriptions of the same brand shared no exact description term. Of those cases with no shared description word, 87.04% still shared a tone. Measuring an exact adjective and measuring sentiment can produce very different views of stability.

Keep the specific claim, the use case and its recurrence. A general positive score can conceal a repeated objection about price or fit; one alarming adjective can exaggerate a transient answer.

Updating a price on a website does not guarantee the next AI answer will use it. Parse followed 706 cases where a page changed a dollar amount, AI had previously quoted the old amount, and a full 60 days of follow-up was available. In 26.5% of cases, an answer showed the new amount within 60 days. In 48.3% of the same cases, at least one later answer still quoted the old amount.

The study detected changes by revisiting and saving web pages, often called crawling. Its clock started when a visit found the changed price, which may have been after the edit was published. The median gap between visits was 15.1 days. The figures describe what happened after detection, not exactly how long each AI platform took after publication.

Citation freshness can also be mistaken for content freshness. The separate page-age study reported a median three-day server-reported modification age, but 11 of its 15 most-cited websites had no measurable citations under the age filters. An HTTP timestamp can describe a server response without proving when an article was written or meaningfully edited.

After correcting a page, check whether later answers to the same question use the new information. A page edit and an updated AI answer are two separate events.

Parse's link-destination study found 16.7% of checked citation links did not land on the cited address. Moved links outnumbered confirmed-dead links by 12 to 1, and 4.8% of links ended on a homepage rather than the cited page. A redirect is therefore a more common inspection problem than a confirmed dead page in that study.

In a separate page check, 0.67% returned a 'not found' or 'gone' response (HTTP 404 or 410). Another access study found that 1.6% returned codes for authentication, blocked access or too many requests (401, 403 or 429). A page that blocks an automated visitor is not necessarily missing for a person.

Different links can also lead to the same page. Tracking tags added after a question mark in a URL often record where a visitor came from. Removing those tags reduced ChatGPT Search's count of distinct cited pages by 9.2% in one Parse study.

To check a citation, follow it and read the passage that is supposed to support the claim. Notice if the link lands on a homepage, a different page or a sign-in screen. Those are different problems from a page that no longer exists.

The GEO paper by Aggarwal and colleagues, published at the KDD research conference in 2024, reported visibility improvements of up to 40% in its experimental AI-answer system. It also found that the effectiveness of optimization methods varied across domains. That is evidence from a particular experiment, not a promised lift on current ChatGPT or Google products.

Google's own Search documentation says its normal SEO foundations remain relevant to AI Overviews and AI Mode, with no special AI markup required for eligibility. It recommends making important content available as text, allowing crawling and keeping structured data consistent with what readers see.

A practical starting point is to make your pages useful for a real buying decision. Explain who a product suits, what it costs, what it works with and what it cannot do. Support comparisons with evidence a reader can inspect. Clear formatting helps people read the page, but a table or FAQ alone is no guarantee of an AI citation.

When someone promises a GEO improvement, ask what improved: a mention, a citation, a recommendation or a sale? Then ask which questions and platforms were tested. A result from an experiment is more informative when you can see the starting point and what was actually changed.

There is no single prompt count that makes every AI visibility report reliable. The questions you choose matter: asking one question repeatedly can miss the different ways people compare products, describe a need or narrow their options.

A Parse study used 4,090 recorded observations across 79 public questions, four deliberately selected markets and two engines. It tested small samples against a reference calculated from that larger, fixed collection of answers.

With ten answers in total, using ten different questions produced an error of 13.1 percentage points. Asking five questions twice produced 15.4 points; asking one question ten times produced 27.3 points. The error measure was root-mean-square error: it measures distance from the reference and gives larger misses more weight. Lower values mean a closer estimate.

The study reused recorded answers; it did not run new conversations for each sample. Its reference describes those selected questions, not all customers or all searches, and it did not weight questions by search popularity. The result supports broader question coverage within that collection. It does not provide a universal minimum or a guaranteed margin of error.

Start with the decisions your customers face. Include questions about the category, prices, alternatives and important requirements. Keep the question set consistent when comparing results over time, and use repeated answers to learn which results are stable.

Start with a question you need to answer. 'Do buyers see our name?' needs a mention measure. 'Are we suggested as a good choice?' needs a recommendation measure. 'Does this lead to business?' needs visits and conversions. No single visibility score answers all three.

For a simple first review, choose one product category and write down the buying questions you want to understand. Check them on the relevant AI products. Note whether your brand appears, whether it is recommended, which pages are cited and whether important details are correct. Look at more than one answer before judging a change.

Then compare the visitors that actually arrive at your website. Do they read useful pages, request a demo, sign up or buy? A small number of valuable visits can matter more than a large number that goes nowhere. Record the visit count alongside the conversion rate so a handful of visits does not look like a dependable trend.

Use the results to choose a concrete improvement: clarify a confusing price, answer a missing compatibility question, correct an outdated description or improve a page buyers already rely on. The research helps you decide what to look for; your own answers and website results show what needs attention.

This guide was checked on October 12, 2026. It combines 115 Parse research reports with 11 primary industry sources. Many Parse studies use some of the same answers, brands or pages, so their sample sizes should not be added together.

Every source retains its own period, product, population and exclusions. ChatGPT is not always ChatGPT Search; Google AI Overviews is not Google AI Mode. A category-level leader is not a single answer's first recommendation. A share of citation links is not a share of answers, and neither is a share of people.

For a presentation or article, cite the originating study alongside the number and say what was counted in the same sentence. For example: 'In Parse's 81,800 matched answer pairs from May–July 2026, ChatGPT Search and Google AI Mode chose different top brands in 63.8% of pairs.' Avoid shortening that to 'AI disagrees 64% of the time.'

The full download contains 389 headline entries exactly as published, with their source links, study descriptions and dates. Some are sample sizes or comparisons rather than separate findings. The second download contains the chart and table values shown here. Both make it easier to find and credit the original research.

Some methods can check only part of the available evidence. The pricing study could compare only a small share of quoted prices. Page-text studies could not read every page. A server's modification date may not be the date an article changed. Those limits are explained beside the relevant findings and in the linked studies.

Company announcements establish large product audiences: OpenAI reported 1.2 billion weekly ChatGPT users in October 2026, and Google reported more than 2.5 billion monthly AI Overview users in May 2026. There is no single count of unique people across these AI products in these sources, and product use includes activities beyond search.

Conductor reported a 1.08% average referral share in its May–September 2025 enterprise sample. That is a dated observed benchmark, not a 2026 census of every website. Your industry, audience and attribution setup determine whether it is a useful comparison. It excludes Google AI Overview and AI Mode visits, which Google Analytics counts as ordinary Google search traffic.

Pew observed fewer traditional-result clicks when summaries appeared, and Ahrefs found lower click-through rate for the first result, compared with an estimate based on other keywords. They use different methods. Neither finding means every site's traffic falls by the headline percentage.

Parse observed a higher first-place rate for cited brands than for named but uncited brands. The relationship was not a causal experiment, and the brand with the most citations frequently was not the first recommendation.

Use a measure tied to your decision and compare like with like: the same engine, set of questions, period and outcome. The studies show enough variation across categories and phrasings that a universal visibility target would conceal more than it explains.

These studies do not establish a universal number. The Parse sampling study supports broader question coverage within its selected set of questions. Define the questions you want the metric to represent before claiming precision.

Explore the studies behind this guide. Choose a topic to narrow the list, sort by title or article update date, and follow any study title to its full findings and methods.

The library includes all 115 Parse reports used for this roundup, including studies of content formats, brand language, different markets and changing answers. The headline column preserves each study's published wording; some headlines describe the sample rather than a percentage.

The full statistics download contains 389 headline entries with their sources and study dates. Article update dates in the table describe when a report was edited, not when its observations were collected.