How to Track GEO: Using Google Search Console as an AI-Visibility Proxy
There is no dashboard for AI citations yet — but Search Console has lagging indicators that tell you whether the work is landing. Here is the measurement routine.
There is no Search Console for AI assistants, so GEO measurement is done in two halves: a manual prompt audit that gives you the direct answer ("are we named?"), and Google Search Console proxies that show whether your AI search visibility signals are moving. Do both monthly, on a fixed list of queries and prompts, and you have something better than a dashboard: a record of whether the work is compounding.
This is the routine we use. It takes about 90 minutes a month and it is the only measurement approach that has survived contact with how these systems actually behave.
Why there is no reliable AI dashboard (yet)
The reasons are structural, not a tooling gap someone is about to close:
- Answers are non-deterministic. The same prompt can produce different products on different days, and personalization, location, and account history all shift the result.
- There is no impression data. When ChatGPT names you, you get no log line, no click report, no referrer you can trust — most assistant traffic arrives as direct or dark social.
- Citations are partially opaque. Assistants cite some sources and not others; some answers are built from training data with no live retrieval at all.
Anyone selling you a precise "AI visibility score" is measuring something adjacent and calling it the thing. What you can measure reliably is whether the inputs are improving — and that is what Search Console shows you.
What Search Console can and cannot tell you
| Signal | What GSC shows | What it means for GEO |
|---|---|---|
| Impressions on question queries | Queries starting with how / what / best / vs | Models and searchers phrase things the same way — rising impressions mean your extractable content is being matched to real questions |
| Brand-name impressions | Searches for your product name | A lagging indicator of being talked about and recommended |
| Which pages earn impressions | Page-level performance report | Tells you which of your assets are citation-shaped |
| Query diversity | Number of distinct queries per page | Broad coverage means the page answers a topic, not a keyword |
| Clicks and CTR | Standard performance data | Weak proxy — assistant traffic often does not show as a click at all |
| AI citations | Nothing | Requires the manual audit below |
| Whether you were named | Nothing | Same — manual audit |
Read that table as the core discipline: Search Console tells you whether the supply of citable content is working, and the manual audit tells you whether it is being consumed.
The four Search Console proxies worth tracking
Open the Performance report, set the range to the last 3 months with a comparison against the previous period, then extract these four numbers:
1. Question-query impressions
Filter queries containing how, what, best, vs, alternative, or for (as in "for freelancers"). Track the total month over month. This is the single best proxy for whether your comparison and FAQ content is matching how people actually phrase needs — to search engines and assistants alike.
2. Brand impressions
Filter queries containing your product name. Brand search volume rises after people encounter you somewhere — a directory listing, an AI answer, a newsletter. It lags by weeks, which makes it a slow but honest confirmation signal.
3. Pages earning impressions on zero clicks
Filter for pages with impressions but near-zero clicks. Pages that get shown but not clicked are often being used to compose an answer rather than to be visited. That is not failure; that is the signature of a page that is being read by something that summarizes.
4. New queries per month
Export queries, diff against last month's export, count the new ones. Growing query diversity means you are being discovered for problems you did not explicitly target — a sign the content is being generalized by retrieval systems.
The manual prompt audit
This is the half that gives you a straight answer. Once a month, on the same day, run the same prompts.
Build a list of 10–15 prompts in three groups:
- Category prompts — "what's the best [your category] tool?", "best [category] for [your audience]"
- Comparison prompts — "[you] vs [competitor]", "[competitor] alternatives"
- Problem prompts — the job your product does, phrased as a question a buyer would actually ask
Run each one in ChatGPT, Claude, Perplexity, and Gemini. Use a fresh chat each time, and turn on web search where the assistant offers it — that is the mode that reflects discoverability.
Record four things per prompt, in a simple sheet:
| Field | Why |
|---|---|
| Were you named? | The headline metric |
| Position | First mention is worth more than fifth |
| Description accuracy | Wrong facts are worse than no mention |
| Sources cited | Your next worklist — the pages the model trusted |
Those cited sources are the highest-value output of the entire exercise. If an assistant answers your category question using three directory pages and one roundup, you now know exactly which four pages to get onto.
Putting it on a monthly cadence
Ninety minutes, once a month, in this order:
- Export Search Console (last 3 months vs. previous) → log the four proxies.
- Run the prompt audit → log named / position / accuracy / sources.
- Diff against last month → note what changed, and pick one explanation.
- Turn citations into work → the pages cited most often become the month's link-building and content targets.
- Refresh one page → retrieval favors recent, updated pages. Pick your best-performing asset and genuinely improve it.
Do that for six months and you will have something almost nobody has: a longitudinal record of how AI assistants talk about your product, and evidence for which channels moved it.
Where the compounding actually comes from
The measurement is only useful if it points at work. In practice, the inputs that move these numbers are consistent:
- Permanent, relevant listings. Structured directory pages are highly retrievable and heavily cited (dofollow directory backlinks).
- Category-relevant authority. Where a mention appears matters more than how many mentions exist (domain rating and AI visibility).
- Extractable pages of your own. A one-sentence answer at the top, a comparison table, a genuine FAQ (the GEO/AIEO guide).
- Consistent facts everywhere. Same name, same one-liner, same category, same URL.
Measurement tells you whether those are landing. It does not replace them.
FAQ
Is there a tool that tracks AI citations automatically? Several tools attempt it, and they are useful as a starting point for the prompt list — but none produces stable, comparable numbers, because assistant answers are non-deterministic. Treat any single "AI visibility score" as directional, and keep your own audit as the source of truth.
How often should I run an AI visibility audit? Monthly. Weekly is too noisy to interpret; quarterly is too slow to catch a problem before it compounds. Same day each month, identical prompt list.
Can I measure GEO purely with Search Console? No. GSC shows whether your citable content is being matched to real questions, which is a strong leading indicator — but it cannot tell you whether an assistant named you. That part is manual.
Does assistant traffic show up in analytics? Sometimes, as referrers from the assistant's domain, but often it does not — many users read the answer and type your brand into a new tab. Watch brand-name search impressions instead; that is where the effect surfaces.
What is the one number to watch? Question-query impressions in Search Console, paired with "were we named?" from the prompt audit. The first is a leading indicator you can move this month; the second is the outcome you are actually optimizing for.
Want the inputs that move these numbers? List your product on aat.ee for a permanent dofollow listing, or see the directory & GEO tiers.