Skip to content
New guide: How to measure AI visibility — and what you can’t measure. Read it →
PRISOL

Methodology

Our methodology — how we assess, improve and measure visibility

"AI visibility" is easy to claim and hard to measure. This page documents how we actually do the work — the inputs, the sampling rules, how we diagnose problems and how we report results — so you can judge it for yourself.

Get an AI Visibility Audit

By · Updated · 3 min read

At a glance

Discover
Goals, buyers, competitors and the questions that matter — agreed in writing.
Analyze
Search data plus repeated AI answer sampling, diagnosed across four layers.
Optimize
Blockers first, owned assets next, third-party work in parallel.
Measure
Same prompts, same rules, on a schedule — reported with method and limits.

Principles

  1. Evidence before opinion. Every recommendation points to data: a crawl finding, a Search Console trend, a sampled answer, a cited source.
  2. Same method, every time. Results are only comparable when prompts, platforms, run counts and coding rules stay constant.
  3. Separate the signals. Rankings, citations, mentions, recommendations and accuracy are tracked separately.
  4. State the limits. Every report says what the numbers can’t show.
  5. No manipulation. No spam links, fake reviews, hidden text or attempts to game AI systems.

Phase 1: Discover

Inputs: your goals, offerings, buyers, competitors, markets and any existing data (Search Console, Bing Webmaster Tools, analytics, CRM).

Activities:

  • Agree the competitive set (four to eight brands).
  • Build the prompt set across category, comparison, problem and brand questions — typically 20–60, phrased the way buyers ask. Sources: search queries, sales and support conversations, community threads, AI assistant testing.
  • Agree platforms and markets: usually Google (including AI Overviews/AI Mode), ChatGPT, Gemini, Perplexity and Copilot.
  • Record your brand fact sheet — the agreed source of truth for entity work.

Output: a scoping document you sign off before measurement starts.

Phase 2: Analyze

Search baseline

Crawl and index health, rendering, crawler access (including AI crawlers and CDN rules), rankings and traffic for the pages that should answer each prompt.

AI answer sampling

Rule Standard
Runs ≥3 per prompt per platform per period
Conditions Logged out where possible; fixed location; mode recorded (search on/off)
Recording Full answer text, cited URLs, date, platform
Coding Mention, recommendation, first-mention position, citations by domain, accuracy of each stated fact, sentiment
Quality check Second-person review of a sample of coded answers

Diagnosis across four layers

Layer Question Typical findings
Access Can crawlers reach and render the content? Blocked bots, JavaScript-only content, unindexed pages
Answers Do our pages answer the prompts clearly? Buried answers, missing pages, vague copy
Entity Are our brand facts consistent and clear? Conflicting descriptions, outdated profiles
Authority Do credible sources describe us well? Absent from cited review sites and roundups

Output: an audit report with evidence for each finding and a prioritised action list.

Phase 3: Optimize

Work is sequenced by impact, confidence and effort:

  1. Blockers — crawler access, indexation, rendering.
  2. Owned assets — answer-first rewrites, missing pages, structured data, brand facts on site and owned profiles.
  3. Third-party presence — profile corrections, reviews, editorial outreach, digital PR (started early because it is slow).

Each action has an owner, a definition of done and the prompts it is expected to affect.

Phase 4: Measure

  • Re-run the same prompt set under the same conditions on a fixed schedule (usually monthly).
  • Report search and AI metrics side by side: rankings, clicks, mention/recommendation/citation share, accuracy issues resolved.
  • Flag changes that fall within normal run-to-run variation.
  • Connect to business signals where possible: AI referral traffic, branded search, self-reported attribution, conversions.
  • Review the prompt set quarterly; new prompts are added as a separate cohort.

What our reports always include

  • Prompt list, platforms, run counts, dates and locations
  • Raw sampled answers and coding
  • What we changed during the period
  • What moved, what didn’t, and our interpretation — clearly labelled as interpretation
  • Limits of the data

What we don’t do

  • Promise placement in any AI platform’s answers
  • Present one-off screenshots as results
  • Present sampled visibility as traffic or market share
  • Use link schemes, fake reviews, hidden text or prompt-injection tactics

FAQ

Methodology: frequently asked questions

How many times do you run each prompt?
At least three times per platform per measurement period, more for high-priority prompts or when results vary a lot. Run counts are always stated in reports.
Do you use tools or manual sampling?
Both, depending on scale. Whatever the tooling, the prompt set, run counts, conditions and coding rules are documented so results are comparable over time.
Can we see the raw data?
Yes. Clients receive the sampled answers and coding, not just summary charts.
What if results don't improve?
We report that plainly, with our view of why, and adjust the plan. We don't change the measurement method to make results look better.

Related to Methodology

AI Visibility Audit

See the method applied to your brand

The AI Visibility Audit runs the Discover and Analyze phases and hands you the evidence.

Prefer email? prithvi@prisol.org