You know your rankings. You know your traffic. But you don't know your AI visibility — whether ChatGPT, Perplexity, Gemini or Google AI Overviews actually cite you — and that gap is now the biggest blind spot in most SEO programs. A site can hold page-one positions across its whole keyword set and appear in zero generated answers.
This is the complete audit playbook. It produces three deliverables: a citation baseline, a competitor ownership map, and a prioritized fix list. Budget: half a day for a 20-page scope, or one automated GEO analysis run.
Why impressions are the right scope — not rankings
The audit starts in Search Console, and it starts with impressions, not clicks. Here's the reasoning: an impression means your page appeared for a query — including queries where the answer was satisfied by an AI Overview and the user never clicked anything. Those queries are exactly where citations are being handed out, and impressions count them while clicks don't.
Scope definition: top 20 pages by impressions over 28 days. That's where authority already exists and where citation effort pays back fastest. (The full reasoning is in our Search Console metrics guide.)
Phase 1 — Build the question set (15 minutes)
For each of the 20 pages, pull its top 5 queries verbatim from Search Console. Don't paraphrase, don't "clean up" — the typos, dialect and fragments are the point. You now have ~100 real questions your audience actually asks.
Three things to log per query, because they'll drive the fix phase:
- Shape — full question / fragment / topic phrase
- Intent — comparison / how-to / troubleshooting / local
- Language/register — English, Arabic, dialect, slang, clinical
Phase 2 — Run the citation test (60–90 minutes)
For each question, ask it as written in each engine. Manual testing is unglamorous but irreplaceable — this is the ground truth every tool calibrates against.
Test matrix per question:
| Engine | How to test | What to record |
|---|---|---|
| Google AI Overviews | Search the query, AI Overview present? | Cited / competitor cited / no citation |
| Google AI Mode | Ask conversationally | Same |
| ChatGPT (browsing on) | Ask fresh session | Same + which sources listed |
| Perplexity | Ask, check source list | Same |
| Gemini | Ask | Same |
Scoring: Cited = +1, Competitor cited = 0, No source shown = 0. Sum per page and per engine. Your citation baseline is now a number you can defend in a meeting — e.g. "12% of our question set gets us cited in ChatGPT; the market leader gets 41%."
Practical notes:
- Fresh sessions per test; logged-in personalization poisons results
- Run the same question 2–3× where the engine varies; record the majority
- Screenshot everything — answers drift weekly, and you'll want the receipts
Phase 3 — Read the ownership map
Cluster your results into three buckets:
Bucket A — Cited and holding (defend). These pages win today. The risk is complacency: engines re-select sources constantly. Keep passages fresh, monitor monthly.
Bucket B — Competitor cited (attack). The question gets answered and someone else is the source. This is your highest-value bucket: demand exists, an answer exists, and the winner is beatable — open the cited page and audit it against the five citable-passage properties. You'll usually find it's just better structured, not better resourced.
Bucket C — No source / training-data answer (watch or seed). Either the query is too niche for a citation, or the engine answered from memory. Re-test after content changes; if a proper answer starts existing on the live web, engines begin citing it.
Phase 4 — Diagnose the losses structurally
Free GEO consultation
A specialist reviews your data and first report with you — within one business day.
Book a free consultationFor Bucket B pages — where you rank but don't get cited — the diagnosis is structural, not authoritative. Run each failing section through the checklist:
- Does the answer appear first under the heading, or after context?
- Does the passage pass the rip-out test (self-contained)?
- Does the heading mirror the question shape you tested?
- Is the language the audience's register or the brand's?
- Are claims specific enough to quote (mechanisms, numbers, entities)?
In our audits, 7 of 10 invisibility cases fail on question-shape mirroring alone: the page answers the topic under a heading that never asks the question. The fix costs an afternoon; the citation often arrives within one re-crawl.
Phase 5 — Fix, then measure the delta
For each prioritized gap, produce the dual-format rewrite — heading block and FAQ pair, scored per the decision framework — and publish the winner. Then:
- Wait for re-crawl (Search Console URL inspection can nudge it)
- Re-run the citation test on exactly those questions
- Update the baseline; report the delta
This before/after delta is the only honest GEO metric. Everything else — "AI-optimized scores", vendor dashboards — is a proxy.
The scale problem, and the two shortcuts
The manual audit is 20 pages × 5 questions × 4–5 engines ≈ 400+ tests, plus ~100 dual-format rewrites and 200 scores for the fix phase. That's a specialist-week, every cycle. Two ways to compress it:
Shortcut 1 — automate the analysis. One GEO run ingests your Search Console property, pulls top pages and real queries, generates every dual-format rewrite, scores all 200 items for citability, and hands you the prioritized publish list. The manual audit above becomes a review session.
Shortcut 2 — make the data conversational. Connect Search Console to Claude or ChatGPT with our free MCP server, and phase 1 becomes a prompt: "my top 20 pages by impressions with their top queries." Re-test triggers are one follow-up question away.
Make it a rhythm
Run the full audit monthly; the citation test on your defended pages weekly. Citation ownership moves on re-crawl cycles, not quarterly plans — the teams that test weekly catch losses while they're cheap to reverse.
Start this week: pull your top 20, run the matrix on your ten highest-impression questions, and find out what the answer engines are saying about you when you're not in the room. If you'd rather have a specialist walk your first baseline with you — book a free GEO consultation and bring the numbers.