Is ChatGPT Search Really 20% the Size of Google? What the Numbers Mean for Marketing

What exactly did OpenAI claim about ChatGPT and Google?

OpenAI's Adam Fry said ChatGPT's search-like query volume was “roughly one-fifth” of Google's query volume in Alex Kantrowitz's interview published November 3, 2026. That company-reported comparison is not verified search market share. The accessible interview does not disclose enough counting, time-period, geography, or Google-denominator detail to reproduce it. For marketers, it is a reason to investigate discovery behavior, not a formula for reallocating budgets.

In the original Big Technology interview, Fry's exact words are: “ChatGPT’s volume of search-like queries is now roughly one-fifth the volume of queries of Google.” The interview was conducted at an OpenAI-hosted dinner in New York on Monday night and published the following day. November 3 is the publication date, not a verified recording timestamp. The public transcript is lightly edited. Fry provides a qualitative description of search-like tasks, but the accessible text stops before his answer to whether each conversation or each question requiring web search counts.

That matters. A search-intent request, a conversation, a question initiating web search, and an internal retrieval are not necessarily the same event. The Verge's November 3 event reporting says Fry described a user query resulting in a web search, but did not specify the source of Google's volume. This is secondary reporting, not a publicly accessible answer in the original transcript.

My starting position is simple: take the statement seriously without pretending its missing methods are available. We need comparable units before turning a provider's scale claim into a market conclusion.

Does 20% of Google's query volume mean 20% search market share?

No. The approximately 20% shorthand comes from Fry's November 3 published one-fifth comparison of ChatGPT search-like queries with Google queries, with the Google denominator undisclosed in the accessible original source. Market share needs a defined market, comparable events, matched coverage, and a common period. None can be reconstructed from that quote alone.

It also does not establish displacement. Someone can research a supplier in ChatGPT and search Google to verify it. A new AI interaction may supplement an existing journey rather than replace a Google query. Conversely, an answer might remove a later search. This statement does not quantify either behavior or distinguish additional demand from substitution.

Do not convert the ratio into Google's lost advertising revenue, website traffic, or an implied daily ChatGPT search count. A different interaction design can produce a different number of events for the same information need. Without the operational definitions, precision would be manufactured rather than earned.

For a CMO, the useful question is not which platform wins a headline. It is where commercially important decisions begin, which evidence buyers inspect, and whether the business can observe the outcomes that follow. Query volume measures activity. AI visibility measures presence. Referral traffic measures a route. None alone measures incremental business value.

How many ChatGPT users actually perform web searches?

The sources describe different populations. Big Technology's November 3 introduction reports 1.2 billion weekly active ChatGPT users, based on Fry's remarks, rising from 1 billion over 60 days. That is company-reported all-product reach, not weekly active Search users, people visiting your website, or a measured buyer audience.

Semrush's April 7, 2026 original analysis says 34.5% of queries enabled web search in February 2026, down from 46% in late 2024. Its methodology uses more than 1 billion lines of U.S. mobile and desktop clickstream data from October 2024 through February 2026. The results represent U.S. sessions found in its 200 million-user clickstream panel; that panel size is not the count of ChatGPT users sampled for this result.

Semrush uses queries or prompts for that statement and search-enabled sessions in neighboring descriptions. Its public methods do not establish equivalence with OpenAI's internal search-like-query measure. The February estimate and November reach claim also have different periods and coverage. Multiplying them would not yield a defensible count of Search users.

The commercial discipline is to preserve the original units. Reach tells you about a product's audience. Search invocation tells you about a feature's observed use under a study's detection rules. Neither identifies your addressable category or the proportion of those interactions that involve a buying decision.

Why can AI activity grow without matching website referrals?

An answer can be useful without sending a visit. A citation might provide reassurance within the interface, while another buyer follows it to check scope or evidence. Those are plausible mechanisms to investigate, not proof that a particular share of buyers behaves that way.

The original working paper, How AI Search Rewrites the Web's Economic Bargain, version 1 dated July 8, 2026, provides a bounded observation. In Table 1, 5.2% of 409,133 ChatGPT conversation sessions produced at least one clean third-party visit in Comscore U.S. desktop activity from October 2024 through July 2025. Its ChatGPT sample includes 56,578 active households. This is a historical session-level clean referral ratio, not today's global all-device Search click-through rate.

The same table reports 31.1% of 61,453,859 Google search queries producing at least one clean third-party visit, from 291,747 Google-querying households in that U.S. desktop period. A conversation session and a search query are different denominators. The paper's default ChatGPT session boundary is one hour, with robustness checks at alternative boundaries. Multiple visits after an event do not turn the clean referral ratio into total clicks divided by sessions.

Use the paper to challenge the assumption that activity automatically sends traffic, not to calculate a universal platform conversion multiplier. It is a working paper, not a demonstrated causal effect of adopting AI. Our analysis of the publisher referral economy explores another part of this problem, but its publisher observations should not be imported as your site's benchmark.

Separate six metrics before asking whether AEO works

I use six measurement categories here. They are an editorial operating framework, not six interchangeable stages with a predictable conversion rate. The GEO measurement methodology guide examines visibility mechanics in greater depth; this framework connects those observations to commercial decisions.

Reach: Record active users by product, geography, and period when a reliable source provides them. Treat external product reach as context, not your brand's exposure. Your category may be a small, unmeasured subset.

Activity: Separate prompts, conversation sessions, web-enabled requests, internal tool calls, and traditional queries. Write the unit beside every number. Do not silently substitute sessions for people or Google searches for ChatGPT conversations.

Answer visibility: Track mentions, citations, recommendation context, and factual accuracy for a fixed buyer-question panel. Preserve engine, mode, date, market, and prompt wording. A brand being named negatively is not a favorable recommendation.

Referral behavior: Measure observable visits, landing pages, engagement, and useful next actions. A missing referrer does not prove an AI origin. A named AI referrer does not establish that AI created new demand.

Commercial outcomes: Connect consented, available acquisition evidence to qualified inquiries, accepted opportunities, pipeline, and revenue in the CRM. Agree definitions with sales. A form submission is not automatically an opportunity, and pipeline is not cash.

Incrementality: Ask whether the intervention added outcomes compared with a credible alternative. Use experiments or carefully designed comparisons where feasible. If controls, sample size, or data quality are insufficient, say the causal contribution remains unresolved rather than assigning every resulting deal to AEO.

Measure Google AI exposure with the current Search Console report

It is now outdated to say Search Console has no dedicated Google AI exposure reporting. Google's Generative AI Search report documentation says the report became globally available on August 31, 2026. It covers link impressions in AI Overviews and AI Mode, excluding Search Labs. This is Google's own reporting scope, not a measure of ChatGPT, every AI answer, or every mention of your business.

Use its page, country, date, and device dimensions to inspect the included Google AI link exposure. Its dates use Pacific Time, so document timezone alignment when comparing exports. Keep ordinary Google Search performance reporting distinct, and do not add overlapping impression datasets as if they represented unique people.

Our Google AI impression reporting guide explains the dedicated report and its boundaries. Combine those exports with a repeatable cross-engine question panel, observable GA4 referrals, and CRM outcomes. Each source answers a different question; disagreement is useful diagnostic information, not something to smooth away.

Assign an owner to an evidence register containing source URLs, export dates, screenshots where appropriate, surface settings, definitions, and known limitations. Keep tracking consent-aware and avoid sending personal information to analytics. A tidy dashboard cannot repair an undefined denominator or an unsupported attribution rule.

What should B2B websites help an AI-informed buyer verify?

A plausible journey is AI-assisted discovery, independent verification, a vendor-page visit, an inquiry, and sales qualification. Buyers can revisit or skip steps, use several surfaces, or contact sales directly. This is a planning model, not an observed universal path.

Make each high-intent page answer the consequential questions: what the service covers, who owns delivery, what evidence supports the claims, which constraints apply, and how to take a suitable next step. Include useful implementation detail, authentic examples, and clear comparisons without inventing testimonials or competitor weaknesses.

Technical accessibility, consistent entity information, and structured data that matches visible content help make a source understandable. They do not guarantee retrieval, citations, or revenue. The same evidence should remain useful when a human arrives to verify it. Optimizing solely for an answer-engine mention risks neglecting the commercial decision your website must support.

If you need to identify gaps, explore our SEO, AEO and GEO service or check your AI visibility. The existing Free AI Visibility Score is a starting point, not a promise of citations or a substitute for a scoped AI Visibility Audit.

Should brands move Google SEO or PPC budgets on this headline?

No budget transfer is justified by this headline alone. Maintain profitable existing search activity while designing a bounded AI discovery test. Define its cost, target intent, markets, duration, acceptance criteria, and stop conditions before publishing new content or buying media.

Compare marginal contribution using account-specific evidence. Acquisition cost, qualification, sales cycle, margin, and retention can matter more than channel volume. Avoid comparing a mature revenue-producing campaign with a new experiment's citation count and calling the latter more efficient.

I would fund a limited measurement capability before recommending a platform migration. If a pilot reveals better qualified demand under a credible design, discuss scaling that demonstrated scope. If it only improves answer presence, report that narrower result. Neither a provider's reach nor a historical referral ratio tells you how much budget belongs in your next campaign.

A 90-day editorial plan, not a promised performance curve

The following is my recommended operating plan. The day ranges are editorial work windows, not measured time-to-results claims. Adapt them to the buying cycle, team capacity, and quality of available evidence.

Days 1 to 30: establish the baseline. Select 30 to 50 commercially important buyer questions as an editorial starting scope, not a statistically guaranteed sample. Document engine settings and repeat runs consistently. Audit crawlability, canonicals, source evidence, service clarity, and entity consistency. Export the relevant Google AI report and ordinary search data separately. Agree sales acceptance and outcome definitions before any content change.

Days 31 to 60: improve a controlled scope. Prioritize two or three high-intent solution clusters as a capacity recommendation. Repair unsupported claims, unclear implementation answers, weak evidence links, and confusing conversion paths. Assign a named reviewer. Maintain comparison pages or cohorts where practical, recording concurrent campaign changes and seasonality. Do not publish a new page for every prompt or manufacture a performance dashboard.

Days 61 to 90: assess and decide. Repeat the question panel under documented conditions. Review answer accuracy, relevant referrals, qualification, and CRM progression. Separate observed movement from causal inference. If the sales cycle has not matured, report leading indicators and pending outcomes. Decide whether to extend, change, stop, or scale the test using its predeclared criteria rather than a favorable screenshot.

For a scoped assessment, discuss the measurement question through our AI marketing strategy service. A roadmap is something to define with your team, not an invented free deliverable attached to this article.

Modi's verdict: win buyer confidence, then prove contribution

The important challenge is not converting a company-reported scale ratio into a bigger marketing statistic. It is making your business discoverable, verifiable, and measurable when discovery and conversion happen in different environments.

AI can influence a decision without sending a visit. That makes attribution harder, not optional. Keep source-qualified visibility evidence, observable routes, commercial outcomes, and causal claims separate. A serious growth team can acknowledge an emerging channel while refusing to abandon profitable search on an unrepeatable comparison. That is the standard I would take into a budget meeting.