AI traffic is tiny. What AI says about your brand isn’t.

Across 100 websites, ChatGPT sent 0.62% as many sessions as Google organic search, Marie Haynes reported in October 2026. In June 2026, Chartbeat reported that AI chatbots account for less than 1% of pageviews across its publisher network. A referral works like a coupon: it counts only the people who bring it back. The people who read an AI answer and stopped there still learned what the assistant said about your company, including what you sell, how you compare with rivals and whether you can be trusted. No dashboard recorded what they read, and it can be wrong. In research by 22 public service media organizations, 45% of the AI news answers they reviewed had at least one significant issue, and in September 2026 a federal court let a defamation claim over Google AI Overviews proceed. Below are the evidence, the reason some claims reach every model while others reach two, and a method for checking your own brand.

AI sends little traffic, and three datasets agree

Marie Haynes, a search consultant, reviewed 12 months of Google Analytics 4 data from 100 websites with more than 145 million sessions (Haynes, October 21, 2026). Sessions from ChatGPT equaled 0.62% of the volume from Google organic search, or 0.73% with the other AI assistants included. “That’s not a typo. 0.62%,” she writes. On the 55 sites that tracked conversions, ChatGPT visitors converted at 13.8% and Google organic visitors at 3.02%, yet Google drove 96.5% of all conversions. She dates the end of ChatGPT’s exponential session growth to August 2025.

Kevin Indig analyzed 12 Google Analytics 4 properties, a small sample that he notes skews toward B2B (Growth Memo, October 19, 2026). From January to September 2026, Direct sessions rose 140.9% while engaged Direct sessions fell 1.0%, so most of the growth came from visits that did not engage. Sessions with identifiable AI traces made up 0.003% of Direct traffic. In his words, “the amount of Direct traffic to longtail pages in my sample driven by AI is not detectable yet.”

Across Chartbeat’s publisher network, ChatGPT referrals grew more than 200% year over year, and AI chatbots still account for less than 1% of total pageviews (Chartbeat, June 2026). Pageviews from Google Search fell 34% between December 2024 and December 2025.

In March 2025 browsing data from 900 US adults, Pew Research Center found that people who saw an AI summary clicked a traditional search result on 8% of visits, against 15% when no summary appeared (Pew Research Center, July 22, 2025). Liz Reid, who leads Google Search, disputes findings like this. She wrote that “total organic click volume from Google Search to websites has been relatively stable year-over-year” and called third-party reports of dramatic declines inaccurate, without publishing figures (Google, August 6, 2025).

Every figure in that table counts clicks or visits, which remain worth watching. None records the answer a person read before deciding whether to click.

Few searchers click past an AI summary

In the same Pew Research Center data, people clicked a link inside the AI summary on 1% of visits to pages with a summary. Users ended their browsing session on 26% of pages with an AI summary, against 16% of pages with only traditional results. For those users the summary was the last page they saw, and it sent no referral to count.

An answer can say that a recall resolved months ago is still open, or that a chief executive who stepped down still runs the company, and a mention count will record only that the brand’s name appeared. We use four terms to keep these apart, and the last is our working definition:

To see whether that exposure changes behavior, Indig ran 632 tests comparing AI visibility with traffic measures. Across the four brands he examined, he could not establish a reliable link between AI visibility and later branded search. We know of no study showing that the exposure is harmless either, so the practical course is to read what the answers say. We made a related case in AI/LLM visibility is not a reputation brief.

Why one claim reaches every model and another reaches two

Our published case study compares three whey protein brands, with names changed, as five AI models described them to buyers: Google Gemini, Anthropic Claude, xAI Grok, OpenAI ChatGPT and Perplexity Sonar. The answers were collected in October 2026. Across the three brands we tracked 25 product claims and 9 buyer concerns.

Every model described all three brands, yet none of our client’s eight claims reached all five. Judged only by whether each model described them, the three brands were level. The claim count separates them:

Our client’s page said “28g protein per serving”, without the serving size, and two of the five models repeated it. A rival’s page said “25g protein per 32g serving”, and all five repeated that, although by weight the two products are level at about 80% and 78% protein. Across the study, claims traveled further when the brand’s own page stated them with their condition. The sentence we proposed for our client carries its condition with it: “28g of protein in a 34.9g serving, about 80% protein by weight.”

Buyer reports of stomach discomfort, with no cause established, reached four of the five models, as many as our client’s best claims. Proof that rested on a document traveled less far: our client’s public test reports reached three models, and a rival’s contents certification reached two.

In client work we have seen competitors revise their wording once or several times a week. The fastest change we have measured took less than 24 hours, from a client publishing our proposed wording to the models repeating it.

We designed the questions and took the measurements ourselves, in one product category over one period. Answers can shift within weeks, and aggregate figures across client matters are not published here.

How often answers are wrong depends on how you count

In research by the European Broadcasting Union and the BBC, 22 public service media organizations in 18 countries, working in 14 languages, reviewed more than 3,000 news answers from ChatGPT, Copilot, Gemini and Perplexity (BBC, October 22, 2025). Of all AI answers, 45% had at least one significant issue and 31% showed serious sourcing problems. Gemini had significant issues in 76% of its responses. The BBC adds that when people see errors, “they blame news providers as well as AI developers.”

When a New Zealand investment platform asked ChatGPT about its own fees, fund choices and competitors, its team found responses that were “confident, fluent, and frequently wrong.” MIT Sloan Management Review opens its October 2026 article with that platform, under a pseudonym (Richter, Lvovich, Van Osch and Richter, October 13, 2026), and the authors interviewed 15 senior marketing and digital leaders. For brands we found accounts like this one but no peer-reviewed error rate.

Error rates differ from one study to the next for four reasons, detailed below:

The BBC and EBU figure counts whole answers, reviewed by the participating broadcasters, and only significant issues. A brand error rate built on other choices cannot be compared with it. Our analysts classify each statement in an answer as a fact (verifiable, and verified against the primary record), a claim (asserted, not verified) or an allegation (an assertion of wrongdoing or fault), then count each class by assistant and by stakeholder question. Each class calls for a different response: a wrong fact needs a correction, an unsupported claim needs evidence, and an allegation needs the record that answers it. Our standards set out how each finding keeps the observed statement, the primary record and our interpretation apart.

A federal court let a defamation claim over AI Overviews proceed

On September 14, 2026, Judge Thomas M. Durkin of the US District Court for the Northern District of Illinois ruled on Google’s motion to dismiss in Keene v. Google LLC, No. 25 C 11431 (memorandum opinion and order). James Keene alleged defamation per se over four AI Overviews. Several stated that he was “serving a life sentence without parole.” Each Overview cited a Wikipedia page, and that page did not contain the false statements.

The court denied the motion for the first two Overviews. It dismissed the third, on substantial truth, and the fourth, for failure to allege publication, both without prejudice. Google had pointed out that the statements were objectively verifiable, and the court wrote that in doing so “Google concedes the very point it is arguing against.” It added: “The mere act of providing a citation to an assertion does not convert that assertion from fact to opinion.”

Keene has so far won only the right to proceed on two of the four Overviews, and no court has found Google liable. Whatever the outcome, the opinion describes a risk any company or person can face: an answer attributed a false statement to a source that did not contain it, and Pew’s 1% figure suggests that few readers open the cited source to check. In such matters we advise a documented account on pages the models read, with each allegation placed next to the record that answers it.

Test the answer many times

Researchers at the University of St. Gallen put the same prompts to ChatGPT, Gemini, Google AI Mode and Perplexity many times (Schulte, Bleeker and Kaufmann, April 2026). Within a 24-hour window, the brands named by repeated runs of one prompt overlapped by 0.33 to 0.48 (Jaccard similarity), depending on the product category. From one day to the next, brand sets overlapped by 0.45 to 0.59, and source sets by 34% to 42%. The standard error of a brand’s detection rate fell below 0.10 at seven runs, and the authors recommend at least seven runs per prompt per day, aggregated over two to four weeks. One author reports a commercial affiliation.

With overlap that low, a screenshot of one answer, flattering or damaging, shows a single run that the next run may contradict. A mention rate is an estimate that needs many runs, and even a stable one says nothing about what each answer claimed. The unit worth tracking over time is the claim, counted across runs and assistants. Follow the source describes how we trace a claim to the pages behind it.

Who is exposed, and what each can do

How to check what AI says about your brand

A company can start this check on its own:

1. List the questions your buyers, investors, regulators and prospective employees ask before they decide about you, in their words.
2. Put each question to each major assistant at least seven times a day for two to four weeks, the minimum the St. Gallen researchers recommend.
3. Save the full text of every answer, with the sources it cites.
4. Sort each statement about you into facts, claims and allegations, and check it against your own published record.
5. Where a true claim is missing its condition, publish the sentence with the condition attached, on the page the models cite.
6. Keep your traffic reports. They still count visits, and this check covers what they miss.

When we do this work for clients, each finding ends with the wording and the page we propose, and the report ranks them in priority order.

What the evidence does not show yet

• Whether AI exposure changes what buyers do. The only direct test cited here found no reliable link to later branded search across the four brands it examined.
• Traffic samples are small: Haynes calls hers small, 45 of her 100 sites tracked no key events, and Indig’s 12 sites skew toward B2B.
• The St. Gallen paper is a preprint, and the traffic analyses are published by their authors rather than in journals.
• We found no peer-reviewed error rate for brands, and any rate depends on the four choices above.
• Google calls third-party reports of dramatic click declines inaccurate and has not published the data behind its position.
• Our whey results come from one study and one period, and models change.
• Keene is a ruling at the pleading stage.

We read these answers for companies and for the people who lead them. The first consultation is free and includes a sample reading of your own website, so you can judge the method on your own pages before you commission anything. Tell us what AI should say about you.

Sources

1. Marie Haynes, GEO is a distraction, October 21, 2026.
2. Chartbeat, Pageviews are down, but AI’s impact is complicated, June 5, 2026.
3. BBC Media Centre, Largest study of its kind shows AI assistants misrepresent news content 45% of the time, October 22, 2025.
4. Keene v. Google LLC, No. 25 C 11431 (N.D. Ill.), memorandum opinion and order, September 14, 2026.
5. Kevin Indig, Growth Memo, Does AI drive more direct traffic?, October 19, 2026.
6. Athena Chapekis and Anna Lieb, Pew Research Center, Google users are less likely to click on links when an AI summary appears in the results, July 22, 2025.
7. Liz Reid, Google, AI in Search is driving more queries and higher quality clicks, August 6, 2025.
8. Shahper Richter, Alla Lvovich, Wietske Van Osch and Alexander Richter, MIT Sloan Management Review, The New Rules of Brand Visibility in AI Search, October 13, 2026.
9. Julius Schulte, Malte Bleeker and Philipp Kaufmann, University of St. Gallen, Don’t Measure Once: Measuring Visibility in AI Search (GEO), arXiv:2604.07585, April 8, 2026.