Skip to content

GW Labs / Method

Method

How we measure what AI platforms say about you

GW Labs (gwlabs.app) measures how ChatGPT, Google's AI, Claude, Grok, Perplexity and three other AI platforms answer the questions your customers ask: whether they name you, where, and whether the facts are right. We ask the same questions under the same conditions every month, so when a number moves, the answers moved, not the way we asked.

The work goes by several names: AI visibility, generative engine optimization (GEO), answer engine optimization (AEO). The measurement underneath should not change with the name. This is ours. It sits behind every GW Labs snapshot, audit, monthly Scorecard and research page.

Quick answers

How does GW Labs measure AI visibility?

We ask AI platforms the questions your customers ask, in fresh private sessions set to your customers' location, and score every answer: named or not, position, and whether the facts are right. Then we read the sources each platform cited.

Which AI platforms does GW Labs test?

The free snapshot covers five: ChatGPT, Google AI Mode, Claude, Grok and Perplexity. The paid audit adds Gemini, Microsoft Copilot and Meta AI, for eight.

Are AI answers the same every time?

No. They change from day to day. One run is a baseline, not a trend. That is why the audit asks the ten most important questions three times on every platform, and why every monthly report compares like with like.

The questions

  • Written in your customer's words, not ours. "Where can I host a private dinner for 40?" beats "best private dining venue."
  • Mostly unbranded: the questions a buyer asks before they know your name. One or two branded questions check the facts.
  • In an audit you see and approve the list before anything runs.
  • Once a question is asked, it is reported. We never drop a question because the answer looked bad.

The platforms and how we run each one

PlatformSnapshotAuditHow we run it
ChatGPTyesyesTemporary Chat, set to Unpersonalized: no memory, custom instructions or chat history.
Google AI ModeyesyesA dedicated test Google account with no search history, no location history and no address. Personalization off.
ClaudeyesyesAnthropic's API, which carries no memory or history, or a claude.ai incognito chat. Each report says which.
GrokyesyesPrivate chat.
PerplexityyesyesIncognito.
GemininoyesThe same test Google account, temporary chat.
Microsoft CopilotnoyesSigned in, personalization off.
Meta AInoyesA new chat for every question, incognito where offered.

Some platforms block signed-out visitors or give them a weaker model. Where that happens we run signed in, with memory and personalization switched off, and the report says so.

Test conditions

  • One fresh session per question. No follow-ups, no "are you sure?"
  • Private mode is not location control. A private window hides your history; the platform still sees where you are. So we set the location separately, to the market your customers search from, check it before each platform, and write it in the report.
  • Re-runs repeat the baseline exactly: same questions, same wording, same platforms and modes, same location. If any of those has to change, we say the comparison is broken instead of reporting the movement.
  • Where a platform shows the model it used, we record it. On API runs the model is held fixed for the life of a client's baseline.
  • A platform we cannot reach cleanly on the day is reported as a gap. It is never filled in.

How many times we ask

  • Free snapshot: 8 to 10 questions, five platforms, one run each. About 40 to 50 answers. A first look.
  • Audit: about 50 questions on eight platforms. The ten that matter most are asked three times on every platform. About 560 answers. Three runs separate a real pattern from day-to-day swing.
  • Monthly plans: about 100 questions a month, under the same conditions as the baseline.

How we score an answer

Four checks on every answer:

  • Named. Are you in the answer at all?
  • Position. First, or how far down? Recommended ("go here") or only listed?
  • Facts. Hours, address, phone, prices, capacity, names, each checked against your own site and the page the platform cited. Wrong is wrong even when you are named first.
  • Sources. Which pages the platform cited. That is the why behind the answer.

They roll up into the numbers we report: the share of answers that name you, the share that name you first, how many platforms name you (for example, 4 of 5), open wrong facts, and who gets named instead.

How we keep the numbers honest

  • Every answer is saved the moment it is read, with the real time stamped automatically, the platform, the mode and the sources cited.
  • Every number in a report is computed by a script from those saved answers. Nothing is typed by hand.
  • Missing data is shown as a gap, never estimated.
  • Every report and research page carries a "How we tested" box: the questions, platforms and modes, location, dates and number of runs.
  • When this method changes, the change is dated at the bottom of this page.

The loop

  1. Test: the questions above, on every platform.
  2. Diagnose: the sources behind each answer, and what is missing or wrong in them.
  3. Correct: through channels you approve.
  4. Re-measure: the same test, every month, reported next to the numbers you are judged on.

This page covers how we measure. What we change for a client, and how, is the paid work.

What this does not tell you

  • It is a sample. Answers vary by person, phrasing, day and account. A customer signed in with years of history may see something different. We test the clean version because it is the one we can repeat.
  • We don't control what AI platforms say. No one does. We measure it and work on the sources they read. We never promise a ranking.
  • We don't buy placements, publish lists ranking ourselves, or write reviews.

Research

Our published research uses this method. Research pages report patterns across five or more organizations and name no client or prospect.

Method version 1. Published November 2026. By Rob Goldstein, founder, GW Labs.
Changes: none yet. Each future change gets a dated line here.

See how the platforms answer the questions your customers ask about you.

Free snapshot, five platforms, one page of findings. No call needed.