In October we asked ChatGPT for an "AEO agency for law firms" twelve times and checked every answer for one of the brands our tracker follows. It was named in none of the twelve answers and cited in none of their sources, an AI citation rate of zero on a question it would very much like to own.
Data as of October 2026: our September and October tracker runs for that brand, 32 prompts asked 12 times each on ChatGPT and on Perplexity, plus a separate 11-prompt run on October 21.
A zero reads like a verdict. Plenty of AI visibility reports present it as one, usually beside a competitor sitting at 9 of 12, and the slide after that is a proposal.
The opposite reading is just as easy to reach. In January 2026, Rand Fishkin, co-founder and CEO of SparkToro, published research in which "600 volunteers ran 12 different prompts through each of the 3 tools a combined 2,961 times." The headline finding: "AIs rarely give the same list of brands or recommendations twice (<1 in 100 times, no matter the question)." If answers churn that hard, a single 0 out of 12 tells you nothing, and neither does the competitor's 9.
His line on vendors was blunter: "any tool that gives a “ranking position in AI” is full of baloney."
What zero out of twelve can and cannot rule out
Start with the arithmetic. It settles more than the argument does.
Twelve answers is a small sample. The standard way to put error bars on a yes or no rate from a small sample is the Wilson score interval, which the NIST Engineering Statistics Handbook says "does not strongly depend upon the value of n and/or p" and was "recommended by Agresti and Coull for virtually all combinations of n and p." Unlike the textbook formula, its lower limit "cannot be negative," which matters when your count is zero.
Run 0 of 12 through it and the 95% range is 0% to 24%. The brand's zero is consistent with a true rate of one answer in four. It does not prove the brand is invisible on that question. It proves it is not common there.
| Times named in 12 answers | 95% range for the true rate |
|---|---|
| 0 of 12 | 0% to 24% |
| 2 of 12 | 5% to 45% |
| 6 of 12 | 25% to 75% |
| 9 of 12 | 47% to 91% |
| 12 of 12 | 76% to 100% |
Read the middle row twice. A firm named in half of twelve answers can truly sit anywhere from one answer in four to three in four, so two firms reported at 6 of 12 and 9 of 12 have overlapping ranges, and a report that ranks one above the other on that basis is ranking noise. The brand's best ChatGPT prompt in October, a question about SEO for companies in its core industry, came in at 9 of 12. The honest way to state that result is somewhere between 47% and 91%.
So Fishkin is right about single prompts. Twelve answers to one question cannot carry a ranking.
Named rate and citation rate measure different things
Before you compare any two numbers in an AI visibility report, find out what was counted. Our tracker records two things for every answer. Named means the brand appears in the text of the answer. Cited means one of the brand's pages appears among the sources the engine returned. Plenty of dashboards blend the two, or report one and label it as the other.
Three October results from Perplexity show how far apart they run:
| Question asked on Perplexity, October 2026 | Named | Cited |
|---|---|---|
| best SEO agency in Houston | 0 of 12 | 12 of 12 |
| best AI SEO agency for getting cited in ChatGPT | 0 of 12 | 12 of 12 |
| best digital marketing agency in Houston | 8 of 12 | 0 of 12 |
On the first two, Perplexity pulled one of the brand's pages into its sources in every one of the twelve answers and then wrote a recommendation that left the brand's name out entirely, so the page was trusted as evidence while the brand behind it was passed over. On the third it named the brand in 8 of 12 answers without citing a single page of its site. The name came from elsewhere.
The Houston row is also one of six changes our tracker flagged as real in October. In September Perplexity named the brand in all twelve answers to that question. In October it named the brand in none, while citing its site in all twelve. The interval on that drop runs from minus 100 to minus 74 points, nowhere near zero. We do not know why it happened.
A citation without a mention is not nothing. It is a content problem you can work on, because the engine already reads the page. A mention without a citation means your name travels through sources you do not control. Both show up as "visibility" on a dashboard that does not split them.
When a citation rate becomes readable
Here is the trap in the other direction. On "AEO agency for law firms," Perplexity named the brand in 2 of 12 answers in September and 5 of 12 in October. A report would print that as a green arrow and a gain of 25 points.
It was noise. The interval on the difference runs from minus 11 to plus 56, and it includes zero.
This is the part of the SparkToro research that gets skipped. In the same piece Fishkin wrote that "visibility % across dozens to hundreds of prompts run multiple times is a reasonable metric." Pooling is what makes the rate usable. Across all 32 prompts at 12 repeats, each engine produced 384 answers in October, and the ranges tighten to something you can act on:
| Engine, October 2026 | Named | Cited | Named rate vs September |
|---|---|---|---|
| ChatGPT | 34 of 384: 9% (6% to 12%) | 30 of 384: 8% (5% to 11%) | -1 point (-6 to +3): noise |
| Perplexity | 154 of 384: 40% (35% to 45%) | 180 of 384: 47% (42% to 52%) | +5 points (-1 to +12): noise |
Even pooled, the instrument has limits. At 32 prompts and 12 repeats it can detect a month over month move of about 6 points on ChatGPT and about 10 points on Perplexity, and anything smaller is indistinguishable from asking the same 32 questions twelve times each the next morning and getting a slightly different count. That is the price of measuring something random honestly. You give up the right to celebrate small wins.
We call a change real only when the 95% interval on the difference excludes zero. Single prompts get the Wilson interval above. Pooled rates come from a Beta-Binomial model per prompt with a Jeffreys prior and 2,000 seeded draws. Neither choice is exotic.
Why the same question gets different sources
Fishkin's volunteers were not seeing a glitch. Profound's AI search volatility analysis, vendor research published in July 2025, found that between June and July the share of cited domains that changed ran from 40.5% on Perplexity to 54.1% on ChatGPT and 59.3% on Google AI Overviews, and 70% to 90% over January to July. SISTRIX reported in April 2026, in a study of 82,619 prompts over 17 weeks that we have only seen summarized, weekly source replacement of 56% for Google's AI answers and 74% for ChatGPT Search.
Our own lawyer data adds a wrinkle. In a check of 100 ChatGPT answers to Houston lawyer questions in late October, the model ran a web search on 85 of them and returned no citation at all on 40. An answer that never searched cannot cite you, so part of every ChatGPT sample is answers that never had a chance.
Perplexity behaved the opposite way, and this is the part we cannot explain yet. In our October run, every one of the 32 questions came back cited in either 0 of 12 answers or 12 of 12. Not one landed in between. ChatGPT, on the same 32 questions, named the brand at an in-between count on eight of them. Our lawyer check showed the same split: across ten repeats of each of ten questions, Perplexity returned exactly one set of sources per question, while ChatGPT produced between 2 and 10 different sets.
Either Perplexity settles on the best sources for a question and holds them, or a cache is serving the same retrieval to every API call that arrives within a short window. All twelve of our repeats each month go out on a single day, which is exactly the condition a cache would exploit. If it is a cache, twelve repeats on one day are closer to one observation than to twelve, and every Perplexity range on this page, the 35% to 45% included, is narrower than it should be. It would also explain why Perplexity's real changes in October were all or nothing: four of its five went the full distance, from 0 of 12 to 12 of 12 or back. The test is to spread the repeats across several days and see whether the source sets start to vary. We have not run it yet. If you have Perplexity API data of your own that answers this, we would like to compare.
Two confounds in these numbers
The ChatGPT figure carries a problem the brand caused itself. From early August until October 15, the firewall on the server that hosted its old site blocked OpenAI's crawlers by user agent: GPTBot, OAI-SearchBot and ChatGPT-User, plus Anthropic's ClaudeBot. Our tracker logs the unblock date as a change in measurement conditions.
Both runs happened inside that block, September's on the 13th and October's on the 4th, eleven days before the unblock. So the minus 1 point is noise, and it is also a comparison between two months in which OpenAI's own crawlers were turned away at the door, which means neither number is a fair read of how ChatGPT treats a site its crawlers can reach.
Perplexity has a confound we did not cause. Every one of its 384 October answers came back with exactly 10 sources. In September, 359 of 384 came back with exactly 20. Same endpoint, same model setting, so the cap changed somewhere between the two runs, and from our side we cannot tell whether Perplexity or the data provider we query through made the change. Half as many source slots per answer is a different instrument, which makes every September to October Perplexity comparison on this page, the five real changes included, a comparison across that switch.
We are publishing both numbers with the caveats attached. One more limit is worth stating plainly: every number here comes from clean API calls with no chat history, while a real person's ChatGPT is shaped by their own history and location. Treat the rates as directional. They are not what any one prospect sees.
What the November run will add
Our November run will be the first ChatGPT month with OpenAI's crawlers able to reach the brand's site for the whole window, on the same 32 prompts so the comparison holds. It will not be clean. The brand moved its whole site to a new platform on October 20, so a November jump would have two suspects, the unblocked crawlers and the rebuilt pages, and a monthly run of 32 prompts has no way to tell which one did the work. The spaced Perplexity repeats should at least settle the cache question.
If you want the lawyer side of this data, our piece on how law firms show up in ChatGPT recommendations breaks down which sources the engines cite for that question. If you are the one sending AI visibility numbers to clients, how agencies should report LLM visibility and the GEO metrics worth tracking cover the reporting side. And if you want your own firm measured this way, with intervals and the named and cited split, that is what our AI visibility audit does.
Matthew (Matt) Bertram helps energy and industrial companies get found, trusted, and chosen as AI reshapes how buyers decide. As owner and CEO of EWR Digital, he leads the agency's AI search and SEO work for oil and gas and other capital-intensive industries; he is also President of ModalPoint. He is also Chief Marketing Officer of the Oil & Gas Global Network (OGGN) and the author of multiple books, including LLM Visibility: A Decision-Grade System for Winning AI-Mediated Discovery and the co-authored Oil & Gas Sales & Marketing: The Energy Growth Playbook for Oil and Gas Leaders. He is a member of the American Petroleum Institute's Houston Chapter and the International Association of Privacy Professionals (IAPP).


