How to Read an AI Visibility Statistic Without Getting Burned

On July 6, Previsible published research putting ChatGPT at 92.4% of trackable large-language-model referral traffic. It is a serious piece of work: 6.77 million LLM-driven sessions across 166 Google Analytics 4 properties, spanning November 2024 to May 2026.

A competing dataset from Lantern puts the same figure at 87.4%.

Five points apart, on the same question, in the same period. Both are real analyses by capable people. Neither is wrong, exactly. And the gap between them is the single most useful thing published about AI search this month, because it tells you something the headline numbers never will: nobody can measure this reliably yet, and the confident decimal points are hiding that.

This is not an argument for cynicism. Real decisions have to be made on imperfect data, and "we can't know anything" is a worse posture than credulity because at least credulity is actionable. The useful skill is narrower and learnable: knowing which questions to ask before a number enters your plan.

Here are six.

---

1. Who produced this, and what do they sell?

Start with the least polite question, because it is the most predictive.

Look at the AI-search statistics circulating this month and note the authorship. The 92.4% traffic share comes from a firm that sells AI-search strategy. The finding that 60% of search journeys now involve AI comes from Brainlabs, an agency that sells GEO services. The finding that google.com is now the second-most-cited domain in AI Mode, up 8.4x comes from Profound, which sells AI-visibility tooling. The finding that AI Mode shows text ads on 29.45% of commercial queries comes from SE Ranking, which sells SEO software.

Every one of these is a vendor measuring the size and urgency of the problem they sell the solution to.

This is not an accusation of dishonesty. These studies are mostly careful, their methodologies are often disclosed, and vendor research is how most marketing measurement has always worked. But it is a structural conflict that no disclaimer removes, and it has a predictable direction: it inflates the addressable problem.

The practical rule: attribute rather than generalize. "Brainlabs says 60% of search journeys involve AI" is a true sentence. "Studies show 60% of search journeys involve AI" is not, and the difference matters when someone senior asks where the number came from.

Sources: Search Engine Land, ChatGPT dominates AI referral traffic (July 6, 2026) · Stacker/KEYT, 60% of search journeys now involve AI (July 23, 2026) · Search Engine Land, Google is AI Mode's No. 2 most-cited domain (July 15, 2026)

---

2. What is the denominator?

This is the question that does the most work, and it is almost never answered in the headline.

"92.4% of trackable LLM referral traffic" is doing enormous labor in that phrase. Trackable, out of what? LLM, meaning which systems?

Compare it to Google's claim on July 18, when SVP of Knowledge and Information Nick Fox told publishers that AI features in Search now send "billions of clicks to websites every week." Billions compared to what? Up or down from last year? Concentrated among which sites? As Search Engine Journal's Matt G. Southern noted at the time, the claim arrived with no methodology, no baseline and no data anyone could check.

A figure without a denominator cannot be falsified, and a figure that cannot be falsified is a talking point rather than evidence. It does not matter whether the underlying reality is favorable — the number itself carries no information.

There is a second-order problem here worth noticing. Site owners cannot independently reconcile Google's aggregate claim against their own data, because Search Console reports impressions rather than clicks for AI surfaces. The one instrument that could verify the claim does not expose the metric the claim is about.

The practical rule: if you cannot state the denominator in a sentence, you cannot use the number in a plan. Write it out longhand before you quote it. "92.4% of the AI referral traffic that our analytics could attribute, from AI systems excluding Google's, across 166 sites that chose to participate" is a much less exciting sentence, and a much more useful one.

Sources: Search Engine Journal, Google Puts A Number On AI Search Clicks, Without The Data (July 18, 2026)

---

3. What has been excluded?

The most consequential fact about the 92.4% figure is not in the figure. It is in the exclusions.

The study omits Google's AI Overviews entirely. And Previsible's own report acknowledges that AI Overviews traffic likely exceeds all standalone LLMs combined.

Read that twice. A widely circulated statistic about AI traffic share leaves out what may be the single largest AI surface on the internet — and says so, in its own footnotes, where almost nobody reads.

This is not the study's fault. AI Overviews referral traffic is genuinely difficult to measure from outside; there is no good public instrumentation for it. The failure is downstream, in how the number travels. By the time "ChatGPT has 92% of AI traffic" reaches a strategy deck, the exclusion has fallen off, and a reader reasonably concludes that non-ChatGPT engines barely matter. That conclusion is not supported by the data and could be badly wrong.

There is a second exclusion stacked on top. The study measures tagged referrals, and by its own accounting 35% to 70% of AI-driven visits arrive untagged, logged in analytics as direct traffic. This is the so-called dark AI traffic problem. The measurable channel is a fraction of the real one, and the fraction is not precisely known — which is why the estimate spans a 35-point range.

The practical rule: ask what is not in the sample, and specifically whether the largest thing is in the sample. An exclusion disclosed in a methodology section is still an exclusion once the chart is on a slide.

---

4. Does an independent measurement agree?

92.4% versus 87.4%. Same question, two datasets, five points apart.

Neither number is "the" number. Both are estimates shaped by sample and method, and the disagreement is not a scandal — it is the expected result of two groups measuring a hard thing in different ways. What matters is what you conclude from it.

The wrong conclusion is to pick the one you prefer, or to average them. The right conclusion is that the precision is illusory and the direction is robust: ChatGPT leads standalone-LLM referral traffic by a wide margin, and nobody knows the exact share.

This is a general test and it is cheap to run. Before you adopt a statistic, spend ten minutes looking for a second measurement of the same phenomenon. If you find one and it agrees closely, your confidence should rise. If you find one and it disagrees, you have learned the error bar. If you cannot find one at all — which is common in this field — then you are holding a single unreplicated vendor measurement, and it should be labeled as such wherever it appears.

---

5. Is the thing being measured even stable?

This one is easy to miss because it is a question about the world, not about the study.

Some things hold still while you measure them. AI citations do not.

Brainlabs' July research, whose headline adoption figures deserve the scepticism in question 1, contains volatility findings that are considerably more useful than its headline:

• Only 23% of AI citations were still live 14 days later.
• AI Overview overlap with Google's traditional top 10 fell from 76% to about 38%.
• Gemini showed only about 15% overlap with the classic top 10.
• Cross-assistant agreement on brand recommendations ran at roughly 45%.

Set aside whether those percentages are exactly right. The direction is corroborated by independent work: researchers reading ChatGPT's own network traffic this month found that identical repeat prompts changed their primary retrieval source about 11.6% of the time, with cited-URL overlap between runs falling roughly 45%.

If a citation has a roughly one-in-four chance of surviving a fortnight, then a single measurement of "AI visibility" is not a measurement of your position. It is a photograph of a moving object, taken with a slow shutter.

This reframes what the metric even is. GEO is not a one-time optimization whose result you check afterwards. It is a maintenance discipline, and the right cadence is continuous rather than quarterly.

Sources: Stacker/KEYT, Brainlabs research on AI in search journeys (July 23, 2026) · Search Engine Land, ChatGPT citations change because of hidden search pipelines (July 8, 2026)

---

6. Can you reproduce any part of it on your own data?

The final question is the only one that fully escapes the vendor problem.

Someone else's average, built on someone else's sample, is at best a prior. Your own instrumented data is imperfect and partial, but it is yours, it describes your actual audience, and nobody sold it to you.

This is less satisfying than a big industry number and considerably more defensible in a room. When a competing vendor claims your category behaves a certain way, the strongest possible response is not a rival study. It is: here is what happened on our properties.

The practical starting point is unglamorous. Instrument AI referral traffic directly in your analytics, accepting that a large share arrives untagged. Watch your own server logs for AI crawler activity. Track your own citation rate across repeated runs rather than single checks. None of that produces a headline. All of it produces a number you can stand behind.

---

An honest steelman for the numbers

This piece has been hard on vendor research, so here is the case in its favor, stated properly.

Directionally, the studies agree, and direction is what most decisions need. Every serious dataset says ChatGPT is the dominant standalone-LLM referrer by a wide margin. Every one says LLM referral traffic is growing quickly. Several say Claude is rising while Perplexity and Copilot fade as standalone referrers. Those signals are cross-corroborated and worth acting on.

A strategist who reads these studies and concludes "AI visibility matters, weight effort toward the leaders, and instrument our own traffic" is reasoning correctly from imperfect evidence. That is what good judgment under uncertainty looks like.

The problem is never the direction. The problem is the decimal point — the moment 92.4% stops being "ChatGPT leads by a lot" and becomes a planning constant that someone builds a budget allocation model on.

---

The pattern is not confined to AI search

It is worth noting that this is not an AI-specific pathology, because the fix is not AI-specific either.

In the same week, an analyst at eMarketer concluded that OpenAI's ChatGPT advertising business is on pace to miss its own revenue target by roughly 90% — eMarketer expects the entire US standalone-chatbot ad market to stay under $1 billion this year and reach about $5.41 billion by 2030, against a reported OpenAI target of $100 billion by 2030.

Separately, two measurement studies showed reported advertising ROI collapsing once the underlying data was corrected. One, attributed to LiveRamp, found identity mismatches cutting measured ROI from $1.50 to $0.43 per dollar spent — turning a profitable campaign into an apparent loss. Another, from CIMM, found a commonly used children's-household targeting dataset was only 42% accurate, implying roughly $590,000 wasted of every $1 million spent against it.

Those are also vendor and industry studies with their own interests, and they deserve the same six questions. But the through-line is hard to miss: across AI search and digital advertising alike, a striking share of the numbers steering budgets are self-reported, modeled, or resting on data nobody has independently checked.

Sources: Adweek, OpenAI's Ad Business Is on Pace to Miss Its Own Forecast By 90%, Analyst Says (July 21, 2026) · PPC Land, Wrong data made a profitable campaign look like a loss (July 22, 2026)

---

Running the checklist on four July statistics

The six questions take about fifteen minutes per statistic. Here is what they return when pointed at four figures that circulated widely this month.

Statistic	Producer / interest	Denominator stated?	Major exclusion	Independent agreement	Verdict
ChatGPT has 92.4% of AI referral traffic	Previsible — sells AI-search strategy	Partly: "trackable LLM referral traffic"	AI Overviews entirely; 35–70% untagged	Contradicted (Lantern: 87.4%)	Directional only. Use as "ChatGPT leads by a wide margin."
AI search sends billions of clicks weekly	Google — the platform being evaluated	No	Unknown	Unverifiable; Search Console exposes impressions, not clicks	Unusable. No denominator, no baseline, no instrument.
60% of search journeys involve AI	Brainlabs — sells GEO services	No	Unknown	None found	Attribute, don't generalize. "Brainlabs says."
Only 23% of AI citations live after 14 days	Brainlabs — same study	Yes, implicitly (citations tracked over 14 days)	Sample size not public	Direction corroborated by independent retrieval-instability research	Most useful number in the batch. Act on it.

Notice what the table does to the ranking. The two figures that led the coverage — 92.4% and 60% — are the two least usable. The one that got almost no attention, the 23% citation-survival rate, is the one that should change how a team works, because it converts GEO from a project into a maintenance discipline.

That inversion is not a coincidence. Headline adoption numbers are the easiest figures in this industry to produce, the hardest to audit, and the most flattering to whoever is selling. Volatility numbers are boring, specific, and operationally expensive to act on. The incentives push the useful findings into paragraph nine.

---

What is safely true right now

Strip out the false precision and a durable set of claims remains. These are corroborated across multiple independent datasets and are safe to build a plan on:

• ChatGPT is the dominant standalone-LLM referrer, by a wide margin. The exact share is unknown and contested; the ranking is not.
• LLM referral traffic is growing quickly. Every serious dataset agrees on the direction and the rough order of magnitude of growth.
• Claude's rise is real and fast enough to have reordered the second tier during 2026.
• Perplexity and Copilot are losing ground as standalone referrers, from earlier peaks.
• AI citations are volatile. Both the 14-day survival research and the run-to-run retrieval findings point the same way, from completely different methods.
• Google's own properties are increasingly cited inside Google's AI surfaces. The 8.4x figure is one vendor's, but the direction has independent corroboration in the behavior of AI Mode.
• A large fraction of AI-driven traffic is invisible to standard analytics, and the size of that fraction is not precisely known.

What is not safe: any share figure treated as a planning constant, any adoption percentage quoted without its author, and any conclusion that a non-ChatGPT engine does not matter — because the measurement excludes the largest surface and cannot support that inference.

---

How to present a contested number internally

The six questions are useless if the answer is always "we can't know." Here is how to use them without becoming the person who blocks every decision.

Lead with the direction, not the figure. "AI assistants are now a material referral channel and ChatGPT dominates it" survives every objection. "AI is 92.4% ChatGPT" does not.

Name the source in the sentence. Not a footnote. If the claim cannot survive being attributed out loud, it should not be in the deck.

State the exclusion where it matters most. If a number omits AI Overviews, say so at the moment someone might allocate budget on it.

Give a range where you have one. 87–92% is more honest than either endpoint and costs nothing in persuasive force. People trust a range.

Separate what you measured from what you read. Your own instrumented data and an industry study are different classes of evidence and should not sit in the same chart without labeling.

Say what would change your mind. "If an independent dataset replicates this, we should increase the allocation" is a stronger position than certainty, and it sets up the next decision.

---

What would actually fix this

The field needs what every maturing measurement discipline eventually gets: a neutral, non-vendor institution publishing AI-citation and AI-referral data with transparent methodology, and some credible way to measure AI Overviews from the outside.

Neither exists yet. Until they do, the numbers steering AI-search strategy will keep being produced by the people selling AI-search strategy, contradicting each other, and omitting the largest surface.

Which makes the most valuable skill in this discipline an unglamorous one: reading the methodology before the budget meeting, and refusing to mistake a confident chart for a settled fact.

We hold ourselves to this too. Any vendor in this category — including us — should be able to tell you what their number excludes, how many runs it is built on, and what would falsify it. If they cannot, the number is not ready to plan around.