You have been asked to justify spending on AI visibility, and the fastest way to do that would be to quote a statistic about how many buyers now start with an assistant. We are not going to give you one, and the reason is the most useful thing in this post: we do not have a defensible market-wide adoption figure, nobody we have found has published one with a method you can check, and repeating an unverifiable number in front of your leadership is a liability you will own personally.
What follows is an evidence audit instead. What you can observe directly, what requires an inference and how big that leap is, what is currently unmeasurable, and how to say all of it to someone who wants a yes or a no.
Why the circulating statistics are hard to trust
There are plenty of figures moving around. They tend to fail for one or more of these reasons.
The measurement base is undefined. "Buyers using AI search" can mean anyone who has ever opened an assistant, anyone who used one in the past month, anyone who used one for a purchase-related question, or anyone whose purchase was influenced by one. These are wildly different populations, and the headline rarely says which one it counted.
The population is not the market. Panels recruited from a vendor's mailing list, a marketing conference audience, or the users of an AI product are not a random sample of buyers in your category. Self-selection points in a predictable direction here.
Self-report is weak evidence for attribution. Asking people whether an assistant influenced a purchase asks them to reconstruct a decision process they did not observe. People under-report influence they did not notice and over-report sources they consider sophisticated.
The publisher has an interest. A large share of the circulating figures come from companies selling AI-visibility products, ourselves included in the category. A vendor-published adoption number is a marketing artifact until a method is attached to it, and we do not treat it as evidence.
Category variance is enormous and usually collapsed. Whatever the true rate is, it is not one rate. Consumer impulse purchases, technical developer tooling, regulated financial products, and enterprise software with a procurement committee do not share a discovery pattern. A single blended number tells you almost nothing about your own market.
So this page states no adoption figure. Any defensible adoption study must carry its sampling frame, question wording, field dates, and raw data; without those, omit the number.
The absence of a trustworthy market figure is not evidence that nothing is happening. It is evidence that the measurement infrastructure has not caught up. Those are different claims, and conflating them is its own error.
What you can observe in your own data today
Your first-party data is weaker than you would like and considerably stronger than any third-party statistic about your market. Three checks are worth running this week.
Check one: referral traffic from assistant domains
Filter your analytics for sessions whose referrer is an assistant or AI-search domain, over a long enough window to see a trend rather than a week. You are looking for whether the count is zero, small and flat, or small and rising.
What this tells you: that some people are following links out of AI answers into your site, and roughly how that is trending. What it does not tell you: total influence, because links followed out of an answer are only one of the ways an answer reaches you. The practical mechanics, including where the referrer disappears, are in the guide on how to detect AI referral traffic in GA4.
Check two: add an open-ended source question to your own forms
On your demo request, trial signup, or contact form, add one optional free-text field: "How did you first hear about us?" Do not offer an AI option in a dropdown, because offering it will inflate it. Let people type.
What this tells you: whether buyers spontaneously mention an assistant, and in what words. Even a handful of unprompted mentions is more informative than a survey statistic about someone else's market. What it does not tell you: prevalence, because form-fillers are a self-selected slice of everyone who considered you, and the field is optional.
Check three: ask your own sales conversations
If you have salespeople, give them one question to ask on discovery calls for a month: "How did you put your shortlist together?" Log the answers verbatim in the CRM.
What this tells you: qualitative texture that no dashboard will produce, including whether buyers mention being given a list of names, and whether yours was on it. What it does not tell you: anything about the buyers who never contacted you, which is precisely the population you most want to know about.
Run all three. The pattern across them is more informative than any one, and the point of the exercise is not to produce a number. It is to establish whether the phenomenon is present in your specific market at all.
What requires an inference, and how large that leap is
Between what you can see and what you would like to conclude sit several inferences of very different sizes. It is worth being explicit about which one you are making.
| Claim | Inference required | Size of the leap |
|---|---|---|
| "Assistants send us some traffic" | None, if you can see the referrals | Small |
| "Assistants named us to people who then searched our brand directly" | Correlating a presence change with branded search movement | Moderate; branded search moves for many reasons |
| "Assistants influenced deals we closed" | Attributing multi-touch influence to an untracked touch | Large; no control group |
| "Being absent from AI answers cost us pipeline" | Counterfactual reasoning about buyers you never met | Very large; effectively unfalsifiable with current data |
| "X% of buyers in our category now use assistants to shortlist" | Generalising from your own funnel to the market | Very large; your funnel is not a sample of your market |
The middle rows are where most confident claims live, and where most of them break. A rise in unattributed direct traffic alongside a rise in your presence in AI answers is consistent with assistants driving discovery. It is also consistent with a competitor's campaign raising category awareness, a seasonal pattern, or a change in how your analytics classifies sessions. Consistent with is not caused by. The full treatment is in the post on what you can and cannot claim causally, and the broader question of untrackable sessions is covered in is dark AI traffic real?.
What nobody can currently measure
Some things are not merely unmeasured. They are unmeasurable with the instruments available, and it is worth knowing which so you stop looking.
-
Answers that named you and produced no click. The impression exists, the analytics record does not. This is the single largest blind spot.
-
Answers that did not name you. You cannot see the shortlists you were left off, and those are the commercially important ones.
-
What any individual user was shown. Personalisation, conversation history, region, and account settings all shape a response. Your measurement of a model is a measurement of the model under your conditions, not under theirs.
-
Real query volume inside assistants. There is no keyword tool for what people actually type into a chat window. Prompt sets are constructed proxies, and honest ones say so.
-
The counterfactual. What would have happened had you been recommended is not observable, which is why anyone selling you a projected revenue figure for AI visibility is selling you a model of a market, not a measurement of one.
Knowing the shape of the blind spot is itself useful. It tells you which claims to refuse when someone else makes them.
A defensible position to take with your leadership
The question you will be asked is binary: is this real, yes or no. The honest answer is not binary, but it can still be decisive. Here is phrasing that holds up.
On the evidence: "We can see a dated number of visits attributed to AI assistants in our analytics. State that measured count and date range. We cannot see how many people were influenced without clicking, and no credible market-wide figure should be substituted for our own observation."
On the direction: "The direction of travel is that recommendations are being assembled before the buyer reaches any of our properties. Whether that is 2% or 30% of our category's buyers, we do not know, and I would not want to present a number we cannot defend."
On the decision: "The proposal is not to bet the budget. It is to start measuring, cheaply, so that in two quarters we have our own baseline rather than someone else's statistic. The cost of measuring is small. The cost of discovering in a year that we have no history to compare against is larger."
On the risk of overclaiming: "If we build a business case on an adoption figure we cannot source, and the figure is wrong, the programme dies and takes credibility with it. I would rather present a small verified number and a plan."
Two conclusions we are explicitly not offering. We are not saying this is hype, because assistants are demonstrably producing shortlists and some buyers are demonstrably arriving from them. We are not saying it is inevitable, because the magnitude is unknown, adoption could plateau, and the interfaces are changing fast enough that today's behaviour is not a safe forecast. The defensible position is that the phenomenon is real, the size is unknown, measuring it is cheap, and claiming to know the size is where people get hurt. For the traffic-side version of the same argument, see zero-click answers.
In Plain English
What this means: buyers are getting recommendations from assistants, and nobody has produced a market-wide adoption number you can safely quote. Your own funnel is a weaker but far more honest source of evidence than any circulating statistic.
Who this helps: anyone being asked to justify or veto spending on AI visibility, and anyone who has been handed a striking percentage and asked to build a plan around it.
When to use it: in the meeting where the question is "is this real yet," before a number goes into a business case.
What to do next: run the three first-party checks, write down what you found including the zeroes, and take the phrasing above into the conversation rather than a statistic you cannot source.
Use the briefing template to record changes in AI answers with methods, dates, and corrections attached.
Primary-source verification
Checked July 28, 2026. The operational and factual guidance above was reviewed against the primary references below. Provider features, prices, policies, interfaces, jurisdictions, and account-specific results can change; recheck the controlling first-party document and your own account or measured data before acting.
- Google Search: AI features and your website - How Google describes eligibility, preview controls, and appearance in its AI search features.
- Google: common crawlers and special-case user agents - Current Googlebot and Google-Extended identities and controls.
- OpenAI: overview of OpenAI crawlers - Current GPTBot, OAI-SearchBot, and ChatGPT-User descriptions and publisher controls.
- Anthropic: web-crawler controls - Current ClaudeBot, Claude-SearchBot, and Claude-User descriptions and controls.
- Perplexity: crawler and user-agent documentation - Current PerplexityBot and Perplexity-User descriptions and controls.
- Apache HTTP Server: mod_log_config - Authoritative Apache access-log fields and formats.
- nginx: HTTP log module - Authoritative nginx access-log configuration and variables.
- Google Analytics: traffic-source dimensions - Current GA4 source, medium, campaign, and default-channel definitions.