How AEO Analyzers measures

How AEO Analyzers measures whether answer engines find, describe and recommend a business — what we ask, which engines, how many times, how we count, and what we refuse to claim. Every number we report is backed by a stored transcript you can re-read.

Three things, measured apart

Found when asked by name. Asked about you by name, does the engine reach your own site? An answer counts only when your domain is among its sources or in its text; an engine that searched, did not find you and said so is a miss, even if its answer then described you from somewhere else.

Described accurately when named. Of the answers that named you, how many stated nothing your own site contradicts? It checks exactly what your site publishes in a form we can read — your brand name, and your founders when your site declares them — and says so beside the number. With no record of your site’s own facts it is reported as unmeasured, never as 100%. It sits beside “found by name”; it does not replace it.

Recommended to new buyers. Asked the questions your buyers ask, without your name in them, does the engine recommend you or someone else by name? Only answers where the engine actually searched the web are scored. Who was named instead is counted from the answers themselves.

The three are never blended into one score. Each has a different cause and a different fix, and a blended number hides which one you have.

The engines, and what they do not cover

Panel v1.1 — ChatGPT, Claude, Perplexity and Gemini — is our series of record. Panel v1.2 adds Grok and starts its own series from its own first date. Every result names the panel that produced it, and no change is ever computed between sweeps on different panels: when the set of engines changes, a shift could be the engines or the day, and we cannot tell which.

Not covered by either panel: Microsoft Copilot, which offers no public way to retrieve its grounded answers, and Google’s AI Overviews and AI Mode, which are features of the search results page rather than something a measurement can call. We do not imitate them. A different set of engines gives a different number — in our own September 2026 measurement, adding one engine to the same runs moved the category figure by 14 points — so the panel is stated on every result.

Test conditions

No personalization. Every question is asked through each engine’s API with web search on — no account, no chat history and no memory. Nothing a past conversation taught an engine can shape the answer, which is a stronger condition than logging out of a browser.

Location. The country is named inside each buyer question rather than left to account or device settings, and it is stated as a test condition on every result. Our own monthly series predates this rule and keeps its questions unchanged, so that series names no country; changing its questions now would break the comparison it exists to make.

The questions

Questions are written the way a buyer asks an assistant, not as keywords. Each one is tagged by the buyer intent it carries — shortlist (who are the options), constraint (options under a hard requirement such as price, size or compliance), alternatives (replacing a named option), comparison (choosing between named options) and problem (describing a situation rather than asking for names) — and the mix is shown on every panel, so you can see whether you are being measured on the questions closest to a purchase. A panel is versioned; when a question changes, the version changes, and the months either side are not compared.

Repetition, N and confidence

Engines answer the same question differently from one run to the next, so one answer proves nothing. Every question is asked several times on every engine, and every figure carries its N with a line saying what N is made of — how many questions, engines and runs, how many answers were scored, and how many were left out and why.

What is left out, and why

• Answers from memory.An engine that answered without searching is reported separately as “answered from memory” and never counted as a zero.
• Failed calls.A run that hit a rate limit or timed out is excluded, not scored as a miss.
• Cut-off answers.An answer stopped by the length cap is not scored.

Reproducibility

Every answer is stored with its question, engine, date and sources. Each stored sweep records which scoring rules it was scored under; when a sweep is reopened under newer rules and a figure moves, the old and new figures and the reason are recorded on it, so the number you saw last month and the number you see now can always be reconciled.

We run this method on our own site every month and publish the result whatever it says: the Honest-Zero ledger.

What we do not claim

We do not know how any engine ranks its sources, and we do not pretend to. We report what the engines actually said when asked. No method guarantees that an engine will recommend a business, and nothing here promises it.

AEO Analyzers (aeoanalyzers.com) was created and is solely maintained by Lindsay Hiebert, founder of PIGENAI LLC. It is unaffiliated with any similarly named browser extension, plugin or tool.