Guide

How S77.AI measures AI search visibility

S77.AI measures a brand's AI visibility by asking AI engines the questions its buyers ask, again and again, and logging every run. Each question gets a citation rate on each engine: the share of valid runs that name the brand or link to it. A baseline comes first, then re-measures at 30, 60 and 90 days.

By · Updated

Why one search proves nothing

AI answers change from one run to the next, even when the same question is asked minutes apart. A screenshot captures one draw. S77.AI asks every question many times and reports how often the brand appears, which is closer to what buyers actually see.

The questions

Each brand gets a fixed set of buyer questions, worded the way buyers ask them and built from search data, the brand's offers and its competitors. Branded questions test how engines describe the brand. Unbranded ones test whether the brand comes up when buyers don't name it.

The set is frozen when the baseline runs. New questions join as a dated addition, reported separately, and no question is ever dropped: cutting the ones a brand loses would flatter the trend.

The engines

How S77.AI reads each engine
EngineHow it is readHow often
Google AI Overviews and AI ModeScripted, signed out, from a fixed New York locationAbout 10 passes a week
ChatGPT, Perplexity and GeminiIn a browser, with a fresh private or temporary chat for every runAt the baseline and at each re-measure
Microsoft CopilotFrom Microsoft's own reports in Bing Webmaster Tools and Clarity, with the brand's accessAt the baseline and at each re-measure

An engine that can't be measured cleanly is reported as not measured. Its number is never estimated.

What counts as a citation

  • Cited: the answer names the brand in its visible text, or links to one of its pages.
  • Look-alikes: names that resemble the brand are screened. A hit next to a known look-alike is flagged for review, not counted.
  • Failed runs: a stalled page, an engine error, a login wall or a CAPTCHA is left out of the count. It is never scored as a miss.
  • No AI answer: when Google shows no AI answer for a search, the run counts as not cited, because the buyer saw no AI answer naming the brand.
  • Personalization: browser runs that show signs of personalization are flagged, and the headline numbers leave them out.

How runs become a rate

On Google's AI answers, a question needs at least 20 valid runs before it gets a state:

  • In: cited in at least half of the valid runs.
  • Contested: cited in more than one run in five, but less than half.
  • Out: cited in one run in five or fewer.

In the browser, a question runs up to three times on each engine at each measurement, so it reads as cited every time, cited sometimes, or not cited. Every rate is computed from logged runs only, so each number traces back to the runs behind it.

Controls

Every round includes a control question the brand should win, such as its own name. If an engine answers the control without citing the brand, the round is flagged for review. A zero is trusted only after the control has passed on that engine.

What a brand gets

Citation rates per engine and per question, never one blended score, because a blended score hides where a brand wins and where it loses. The rates are the baseline of the GEO audit, which puts the fixes in priority order, and the re-measures at 30, 60 and 90 days show what moved.

S77.AI runs this method on its own site and on its sister companies' sites, about 10 times a week.

The fixed point it measures from

Signed-out answers from New York are the fixed point every run is measured from. Answers shaped by a buyer's own history, account or location vary around it, and only moves larger than the run-to-run noise are read as change.

Sources

  1. Google Search Central: AI features and your website
  2. Aggarwal et al., "GEO: Generative Engine Optimization", KDD 2024

Keep reading

Talk to S77.AI

Have a question about this for your brand? Write to info@s77.ai or see the contact page.