How to Measure AI Search Visibility (When There Are No Rankings)

The awkward part of AI search is that the thing you are optimising for is largely invisible in your analytics. Someone asks ChatGPT for a recommendation, reads the answer, and either never clicks or arrives three days later via a branded Google search. Your dashboard records "direct traffic" and shrugs.

You can still measure this properly. It just requires building the instrument yourself.

Step 1: Build a fixed question set

Write 20 to 50 questions a real buyer would ask, spread across the funnel:

• Category discovery — "what tools do X do"
• Comparison — "X vs Y for Z"
• Constraint-loaded — "best X for small teams in the UK"
• Objection — "is X worth the money"
• Branded — "what is [your brand]"

Freeze this set. The value comes entirely from asking the same questions on the same schedule. Changing the set resets your history.

Step 2: Run them on a schedule

Monthly is the right cadence for most businesses; weekly if you are running an active campaign. Run each question on every engine that matters to you — ChatGPT, Perplexity, Google AI Overviews, Gemini, Copilot — in a clean session with personalisation and memory off, since a logged-in session skews toward what you have already discussed.

Step 3: Record four metrics

| Metric | Definition | Why it matters | |---|---|---| | Mention rate | % of questions where your brand appears at all | The broadest signal of presence | | Citation rate | % where your domain is a listed source | Measures your owned content directly | | Source mix | Which third-party domains are cited instead | Your placement target list, handed to you | | Sentiment and framing | How you are characterised — leader, budget option, caveated | Determines whether a mention converts |

Source mix is the most actionable column and the one most teams forget to log. It tells you precisely which publishers the engines already trust in your category. Those are the outlets worth appearing in.

Step 4: Add the indirect signals

Alongside the manual test set, watch:

• Branded search volume in Search Console. AI-influenced buyers frequently search your name afterwards. A rising branded curve with flat non-branded is a strong AI-visibility signal.
• Referral traffic from chatgpt.com, perplexity.ai, copilot.microsoft.com and gemini.google.com. Low volume but unusually high intent — segment it and measure conversion rate separately.
• Direct traffic to deep pages. Nobody types a long URL from memory. A rise in direct hits on interior pages often means an unlinked mention somewhere.
• A self-reported attribution field on your signup or contact form: "How did you hear about us?" It is the crudest instrument available and frequently the most accurate one you have.

Step 5: Interpret the movement honestly

Three cautions:

Answers are non-deterministic. The same question can produce different sources on consecutive runs. Never draw conclusions from a single test — use the rate across your whole question set.

Lag is real. Editorial placements typically take four to twelve weeks to show up in retrieval, and longer to influence anything model-internal.

Zero-click is the norm. A high mention rate with flat referral traffic is not failure. Check branded search and pipeline quality before concluding otherwise.

A simple scorecard

Track these five numbers monthly and you have a defensible picture:

1. Mention rate across the fixed question set
2. Citation rate for your own domain
3. Number of distinct third-party publishers citing you
4. Branded search impressions (Search Console, 28-day rolling)
5. Self-reported AI attribution on inbound leads

Plot them over six months. Individually each is noisy. Together they move in the same direction when the work is landing — and stay flat, unambiguously, when it is not.

Why this beats waiting for a tool

Tools for this are arriving quickly and some are good. But every one of them runs the same fundamental method: a fixed question set, executed on a schedule, with results logged. Building it yourself for one month costs a few hours, teaches you exactly which publishers dominate your category, and gives you a baseline no vendor can hand you retroactively.

Start the log now. Six months of your own data is worth more than any benchmark report.