METHOD — HOW THE MEASURING IS DONE
How I measure what AI says about a business
When somebody asks AI “who are the best Golden Retriever breeders in Oregon?” or “which CPA firms in Portland handle construction companies?”, the AI writes an answer and, more often than not, cites a handful of web pages. Whether your business is named, and whether your pages are among those cited, is what I mean by AI visibility. Here is exactly how I measure it, so you can check my work.
Three named engines
I ask the same questions of three AI engines, and I name them in every report: Claude (Anthropic), ChatGPT (OpenAI) and Perplexity. They do not agree with each other. In the September 2026 Breed Study, Perplexity cited an average of 39.7 sources per answer, Claude 34.4, and ChatGPT 10.5 — and the sites they leaned on were different. A count measured on one engine tells you about that engine. I never report a number without saying which engines it came from.
Fixed questions, written down in advance
Every check uses a fixed list of questions, worded the way a real buyer would type them, agreed before the run and kept unchanged from run to run. That is what lets us compare this month with last month. If a question changes, the report says so and the old figure is retired.
Repeated runs and ranges, not a single number
AI answers change. Ask the same question twice on the same day and you can get a different set of sources. Published measurements of this churn run from roughly a tenth to almost half of the cited sources changing between runs, depending on the engine and the time gap. So the check asks each question more than once and reports a range — “named in 4 to 6 of 9 answers” — rather than a single count that happened to land on one run. Any report that gives you one run and calls it your visibility is giving you a coin flip.
Dates on everything
Every figure carries the date it was measured. AI answers drift over weeks, so a number from July does not describe September. I re-run before I reuse.
The runs where nothing came back are kept
Sometimes an engine answers without searching the web at all, or answers without citing anything. Those runs are kept and counted. In the Breed Study, 39 of 735 answers came back with no source list at all — 37 of them from ChatGPT, 2 from Claude — and they stay in the dataset. Throwing them out would make every “share of citations” figure look better than it is.
The evidence ladder I grade claims by
Not all evidence about AI visibility is equal. Here is the ladder, from strongest to weakest. I tell you where each of my own claims sits on it.
- Field trial. A real change to a real site, with a control, and logs of what happened to actual traffic or enquiries. Rare, expensive, and the only thing that proves a business result.
- Live engines with repetition. Real Claude, ChatGPT and Perplexity, the same questions asked several ways, several times, on dated runs. This is where the Visibility Check and the Breed Study sit.
- Hand-fed engine. A real engine, but with the pages handed to it directly instead of it finding them. It tells you what the engine does with a page once it has it; it tells you nothing about whether the engine would ever have found the page.
- Reproducible pipeline. A research setup that imitates an engine — a search step feeding a language model — that anyone can re-run. Useful for comparing tactics; not a real engine.
- Fixed-context lab test. A language model given a fixed set of five pages and asked to answer. Whatever it measures, it cannot measure whether your page gets retrieved in the first place.
About the “40%” you may have been quoted
Most sales decks for “AI optimisation” quote a figure of roughly 40% more visibility, and it traces to one peer-reviewed paper (Princeton and IIT Delhi, 2024). What that study measured was this: a language model was handed the top five Google results for a question, one of those five pages was rewritten to add quotations, and the rewritten page’s share of the answer’s words rose from about 19 to about 27 — a relative gain in the region of 40%, on one metric, for one tactic, with the page already among the five sources the model was given. It did not measure whether a page gets found, indexed or retrieved, and the 2026 survey of the field describes reading it as a general visibility lift as rejected. It is the bottom rung of the ladder above. I do not quote it, and I would be wary of anyone who does.
What I do claim
Being retrieved is the gate. A page that is already in the answer can raise its share by carrying verifiable, dated, attributed facts. And the only honest measurement is the one above: named engines, fixed questions, repeated dated runs, ranges, nulls kept. That protocol is the product; the number is just its output.
What the check does with your kennel
Puppy buyers now ask AI where to find a well-bred puppy, and the AI answers with a short list and a few links. I check whether your program is on that list. I ask three AI engines — Claude, ChatGPT and Perplexity — the same five questions a buyer in your state would ask, more than once, and I write down the date each time. I tell you how many of those answers named your program and how many linked to your pages, as a range, because the answer changes from day to day and a single run can mislead in either direction. If an engine answers without naming anyone, I count that too. I do not promise a percentage. I show you the answers, with dates, and you can run the same questions yourself and see the same thing.
One thing the study taught me, worth knowing before you spend a dollar: in the September run across twelve states, ten websites accounted for 1,352 of the 3,884 links the AI cited, and one puppy marketplace alone was linked 315 times in 84 of 180 answers. The AI mostly sends buyers to marketplaces and registries, not to breeders. Being one of the exceptions is the whole job.
Sources
- Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande. GEO: Generative Engine Optimization. KDD 2024 (first arXiv version November 2023). The origin of the “40%” figure.
- Martinez, O. A Critical Survey of Generative Engine Optimization. arXiv preprint, July 2026. The evidence ladder, the pipeline model and the minimum measurement protocol summarised above.
- Liu, Zhang and Liang. Evaluating Verifiability in Generative Search Engines. EMNLP 2023 (findings). Citation is not the same as support.
- Puerto et al. C-SEO Bench. NeurIPS 2025 Datasets and Benchmarks track. Generic tactics generalise badly across domains.
- Kirsten et al., ACL 2026, on Google AI Overviews (4,706 queries, US and Germany): overlap and re-run churn figures.
- Schulte et al., 2026 preprint, four engines over 45 days: day-to-day source churn and the case for seven to eight repetitions per prompt.
- Direct-Me. The Breed Study, September 2026 — 180 answers, 12 states, 5 questions, 3 engines (run 2 September 2026), plus a 27-state extension run 14–16 September 2026 (555 answers). Dataset published as CSV.
Questions about the method
Which AI engines does the check use?
Three, named in every report: Claude (Anthropic), ChatGPT (OpenAI) and Perplexity. They do not agree with each other, so a count measured on one engine tells you about that engine only.
Why a range instead of one number?
Because AI answers change. Ask the same question twice on the same day and you can get a different set of sources. So each question is asked more than once and the result is reported as a range with a date, rather than a single count that happened to land on one run.
What about the “40% more visible” figure?
It traces to one 2024 paper in which a language model was handed five pages, one page was rewritten to add quotations, and that page’s share of the answer rose from about 19 to about 27. It did not measure whether a page gets found, indexed or retrieved. I do not quote it.
Can I check the work myself?
Yes. The questions are fixed and written down in advance, every figure carries the date it was measured, and you get the transcripts. You can put the same questions to the same engines and see the same thing.