AI Search and SEO Statistics 2026: Evidence and Limits

Google AI Overviews: Prevalence and Impact

A prevalence statistic answers a narrow question: how often did a specified search sample produce an AI Overview at a specified time? It does not directly measure audience size, unique people, clicks or lost revenue. A keyword database and a panel of real browsing journeys can both produce useful findings without describing the same population. Their percentages should remain separate.

Semrush’s AI Overviews study, updated December 15, 2025, reports 15.69% of sampled keywords showing AI Overviews in November 2025. Its prevalence analysis covered more than 10 million database keywords across January–November 2025. The publisher’s methodology does not provide a complete geographic and device breakdown for that prevalence series. Treat this as a historical database observation, not a probability for every worldwide Google search or a measured 2026 rate.

The same edition reports 6.49% in January and nearly 25% in July 2025. That non-monotonic pattern matters: a rising early observation cannot justify extrapolating continuous growth. A later publisher edition may also use a revised dataset. Keep the edition with the statistic rather than silently extending an older chart with new observations.

Conductor’s 2026 AEO / GEO Benchmarks Report provides a separate historical observation: 25.11% of approximately 21.9 million unique Google searches from its traditional search index generated an AI Overview. Its stated collection window is September 15–October 12, 2025, in the United States. The report page, checked October 4, 2026, does not show a publication date. The earlier article incorrectly assigned this sample to Q1 2026. It should not be blended with the keyword series above or treated as a population-weighted global rate.

The original article combined incompatible research into a broad current range. This revision does not reproduce that range. It also withdraws the claim that virtually all informational searches trigger a summary. The share of AI summaries associated with informational intent is a different conditional probability from the share of informational searches that receive a summary.

For a site-level assessment, build a fixed query set from commercially relevant topics. Record country, language, device, date and whether the result actually displays the feature. Separate branded navigation, research and purchase intent before comparing periods. Keep an unchanged comparison panel and document additions; otherwise a shift toward question queries can look like a platform change.

A useful result is an exposure map of your own content: which topics encounter the feature, which landing pages depend on those topics, and which pages still attract useful visits. This prioritizes investigation without converting an external study into a forecast of your traffic. It also makes unexpected movements reviewable, because the underlying queries remain available.

Google AI Mode: Dated Adoption Evidence

Alphabet’s July 23, 2025 earnings call reported more than 100 million monthly active AI Mode users in the United States and India. This is a company-reported historical audience measure. It is not a daily-active-user figure, a global count, an independently audited audience estimate or a fresh observation for October 2026.

The disclosure does not provide enough detail to reconstruct the active-user threshold or account deduplication. Preserve the company’s metric name. Do not convert monthly activity into daily activity by dividing by days, because the same user can return repeatedly and activity is uneven. Likewise, a user count cannot be converted into search volume without a measured frequency distribution.

The original daily-user estimate and its associated growth multiplier have been removed. The defensible adoption observation still establishes that this was a substantial product audience at that earlier date. It does not establish how frequently those users clicked out, whether their visits replaced traditional search, or what share could become customers of a particular publisher.

For planning, distinguish the surface a person uses from the outcome your business wants. A monthly audience disclosure can justify including a product in research. It cannot set a traffic target. A destination site needs its own attributable sessions, landing pages and accepted enquiries to assess commercial contribution.

When a new adoption release appears, compare definitions before calculating growth. Check whether it extends into additional countries, changes the active period, adds another application or includes a different account population. A larger later number may reflect genuine adoption, broader coverage, a metric revision or some combination. Where the disclosure cannot separate these, explain the limitation instead of assigning the whole difference to growth.

Google’s documentation also says AI Mode and AI Overviews can use different models and techniques. That supports testing both surfaces independently; it does not supply a universal citation-overlap percentage. Keep a result capture for each surface and avoid treating an answer observed in one as proof of visibility in the other.

Primary disclosure: Alphabet earnings-call transcript, July 23, 2025.

ChatGPT Search: Define the CTR Denominator

This article does not retain an overall ChatGPT search click-through-rate benchmark. The earlier figures were attributed to unspecified partner reports without a reproducible public sample, measurement period or denominator. Describing them as estimated would not repair those missing foundations. Their absence here is an evidence limit, not a claim that no such dataset exists.

A click-through rate needs both counted clicks and counted opportunities to click. Possible denominators include all prompts, prompts that invoke search, answers containing a citation, displayed citation links and conversations. Each answers a different question. A comparison between all prompts on one product and search-result impressions on another can create an impressive ratio with little practical meaning.

A destination-site analytics system generally observes arrivals, not every answer that displayed its URL. Therefore referral sessions alone cannot establish platform-wide citation CTR. Some arrivals may lack a usable referrer, while a recorded session may contain several page views. Keep sessions, users and page views separate when presenting the result.

For a controlled visibility study, specify the prompt set, product, account state, locale, observation dates and handling of repeated answers. Capture the actual output and destination URL. Decide in advance whether a brand mention without a link counts as visibility. A public experiment should publish enough of this protocol for someone else to understand the denominator.

Fictional measurement example: suppose a controlled test records 2,000 eligible answer displays and 30 outbound clicks. The display-based CTR is 30 ÷ 2,000 = 1.5%. If the same test contains 10,000 total prompts, clicks per prompt are 30 ÷ 10,000 = 0.3%. These invented inputs illustrate denominator choice only; neither result is a ChatGPT benchmark or a Digital Applied measurement.

For a business report, label the measure with that denominator in the chart title. Pair the referral trend with qualified actions after arrival, and keep the visibility experiment separate from production analytics. A citation can be useful evidence of discovery even when it produces no observed visit, but assigning it revenue requires an additional attribution method.

Perplexity: Users, Queries and Referrals

The earlier Perplexity section mixed monthly users, website visits, query volume, recurring revenue and valuation into a growth narrative. This revision withdraws those numerical claims because a matching, dated primary measurement was not established for each one. Public reporting may exist; a repeated number in a roundup is not sufficient to reconstruct its population and collection method.

A query is an interaction, a visit is a browsing event, and a monthly user is an audience measure under the publisher’s definition. They cannot be substituted for one another. A product can process more queries because existing users ask more follow-up questions even if its active audience remains unchanged. A website visit estimate can omit activity in other interfaces.

Referral share is a different measure again. It describes traffic arriving at the destinations covered by a particular measurement system. It depends on both how much a platform is used and how often that usage sends people to those destinations. A service can grow in absolute referrals while losing share if other sources grow faster.

Fictional referral example: a site receives 300 sessions from one AI source out of 1,000 measured AI referrals in an initial period, then 400 out of 2,000 in the next. Its source-specific sessions increase by about 33.3%, while its share falls from 30% to 20%. All inputs are invented. The example explains why a falling share alone cannot establish shrinking usage.

To evaluate Perplexity for your own audience, use consistent referral classification and inspect the pages that receive visits. Check whether the visitors complete the same meaningful actions used for other channels. Preserve the reporting window and any changes in consent, tracking or domain classification. Where the sample is sparse, show the underlying counts rather than an unstable league table.

A research-oriented product may be worth testing for a specialized audience even when its global position is uncertain. Choose test topics from actual customer questions and preserve the answers you inspect. The useful outcome is evidence about your information and your audience, not a speculative ranking of entire platforms built from incompatible measures.

AI Search Advertising: Observations and Forecasts

The previous advertising section presented a market-spend forecast, an ad-presence rate and a revenue comparison as a single growth story. Those measures require different evidence. This revision removes the unsupported series rather than relabeling it as current market data. It does not supply a replacement spend total without a traceable forecast edition and category definition.

An ad-presence rate should identify the eligible search sample and the location counted: within an answer, above it or elsewhere on the results page. Seeing an ad somewhere on a page containing an AI Overview does not establish that the ad was inserted inside the generated answer. A screenshot proves a placement existed in that captured setting; it does not establish a prevalence rate.

A forecast must state its publication date, target period, geography, currency and definition of AI search advertising. It should explain how the category relates to existing search spend. If the forecast is later revised, retain the old edition only as an earlier expectation. Forecast growth is not observed spending, and a prediction for a completed year still needs actual results before it can become a historical fact.

Platform revenue is also not an advertiser efficiency benchmark. A rise in aggregate search revenue cannot isolate the incremental effect of an AI placement, demonstrate that no other placement lost revenue, or establish campaign profitability. That requires a design capable of separating the treatment from other changes in demand, auction conditions and inventory.

Fictional campaign example: assume $6,000 in spend produces 120 leads, of which 24 meet a sales qualification rule. Cost per lead is $50, and cost per qualified lead is $250. If qualification falls to 12 with spend and lead count unchanged, the latter becomes $500 while the headline lead cost stays $50. These are invented campaign inputs, not market benchmarks.

Use this distinction when assessing an emerging placement. Agree the qualification rule, conversion window and commercial objective before launch. Where placement-level reporting is unavailable, describe the result as a broader campaign outcome. Avoid assigning all improvement to the AI surface merely because it was introduced during the same period.

Zero-Click Search: What the Evidence Measures

Pew Research Center’s July 22, 2025 analysis used browsing from 900 U.S. adults during March 2025. It reconstructed Google results on April 7–17, so the captured summaries could differ from what participants originally saw. The dataset contained 68,879 unique searches. These are historical panel observations, not a global census or a controlled experiment.

In that study, traditional-result links received clicks on 8% of visits with an AI summary, versus 15% without one. Links within the summary received clicks on 1% of visits with a summary. Pew matched next-visited URLs against first-page results and up to the first three summary sources; a link appearing in both was assigned to the summary. The figures do not measure every possible later click.

The traditional-link difference is 7 percentage points. Relative to the group without a summary, the rate was about 46.7% lower: (15 − 8) ÷ 15. This is an arithmetic comparison between observed groups, not a causal estimate of traffic lost because an AI feature appeared. Query intent, result composition and the reconstruction timing remain important qualifications.

Do not subtract the traditional-link percentage from a whole and call the remainder a universal zero-click rate. The missing category can include summary clicks, continued searching, other navigation and session endings. A measure of no external click also differs from a measure of no traditional-result click. Those definitions must be visible before comparing studies.

For your site, begin with a query and landing-page analysis. Separate declining impressions from declining click-through rates, and account for changes in ranking and search demand. Annotate releases, indexing problems, tracking changes and seasonality. A simultaneous feature rollout is a hypothesis to investigate, not sufficient attribution on its own.

The practical decision is where to invest next. A page with valuable demand and weaker click-through may need a better reason to visit: original evidence, a useful calculation, a detailed comparison or a task the visitor can complete. Measure the resulting qualified actions. A traffic decline does not automatically mean commercial performance declined by the same proportion.

Primary study: Pew Research Center: Google users and AI summaries, July 22, 2025.

AI Search Market Share: Keep the Markets Separate

There is no single market-share denominator covering chatbot activity, traditional search, AI-assisted search and referral traffic. A chart must define the market before it names a leader. Share of visits to a selected group of domains is not share of all search queries, and neither is share of people who use AI to make a purchase decision.

This revision withdraws the earlier platform-share ranges and their blended bars. Averaging estimates from different providers does not make a consensus dataset. Their panels may cover different countries, devices, applications and time windows. Even if each estimate is valid within its own scope, a combined ranking can imply comparability that the research does not support.

When using a third-party market series, preserve the provider, edition, geography, device coverage and included platform list. Check whether visits to embedded experiences are counted under a host product. Document whether the share is based on visits, unique visitors, search queries or referrals. Only calculate change between periods when the measurement rules remain comparable.

For an owned-site report, use a narrower title such as share of classified AI referral sessions to this website. That describes a real measurement boundary. Include the total session count and note sources that cannot be classified reliably. It does not claim to represent an entire industry, a country or the world.

A sensible investment decision can combine those narrow observations with customer research. Ask which sources buyers used, inspect the information they encountered and compare the resulting enquiries. Treat self-reported discovery as another imperfect signal, since people may forget or combine several touchpoints. Agreement between independent signals is more useful than false precision in a global share chart.

Finally, separate platform concentration from business dependence. Your site may depend heavily on a source that represents a modest part of someone else’s market panel. Conversely, a widely used service may send little useful traffic to your topic. Set monitoring and content priorities around the exposure you can observe and the decisions customers need to make.

Citation, Visibility and Useful Optimization

A citation is evidence that a particular answer included a source. A brand mention, linked citation, referral visit and accepted enquiry are distinct outcomes. Record them separately. A content-format distribution among cited pages cannot show that changing your page to that format will cause a citation; the result can reflect the topics and prompts sampled.

The original content-type citation percentages, Reddit share and ranking-overlap trend are withdrawn. Their primary dataset, collection dates and coding rules were not established sufficiently for reuse here. This also removes the implied promise that a particular page format or rank guarantees an AI citation.

Google Search Central’s AI-features guidance, updated December 10, 2025 and checked for this revision, says supporting pages must be indexed and eligible for a Search snippet. It does not require special AI markup or an additional AI text file, and eligibility does not guarantee selection. Its documented reporting places AI-feature traffic within Search Console’s Web performance data.

That guidance supports a practical technical review: confirm that important pages can be crawled, that canonical destinations are coherent and that the visible content clearly answers the intended question. Make the evidence easy for a human reader to inspect. If a statistic has a date, population or limitation, put it near the number rather than relying on a distant generic sources list.

For a citation experiment, select prompts before examining the results, retain unsuccessful observations and distinguish repeatability from a single appearance. Track whether the answer represents the source accurately. A link next to an unsupported or misquoted claim is not automatically valuable visibility. Review the context as well as the presence of a URL.

Combine this work with conventional SEO and conversion analysis. Prioritize pages that contain information your audience needs and that your organization can substantiate. Improve the underlying evidence when it is weak; cosmetic changes cannot fix an unsupported claim. When results improve, state what changed and what the experiment could actually measure, without promising a universal citation uplift.

Primary guidance: Google Search Central: AI features and your website. For related planning, see our AI search measurement overview; retain the date and scope of any figure used.

Frequently Asked Questions