How do you measure whether GEO work is actually working?
GEO results can be measured in four ways: traffic from AI sources tracked separately from other traffic, Search Console impressions and positions, a monthly citation check against the same list of questions, and technical check results before and after the work. Three things cannot be measured: how many people read the answer without clicking, when a model will refresh its data, and why in one specific case it cited one source rather than another.
A service with no measurement method is the seller's opinion about their own work. With GEO that is especially convenient for them: part of the result genuinely is invisible, and anything can be explained with that.
But four things can be measured precisely, and those four are enough.
1. Traffic from AI sources
AI assistants send visitors with their own referrer, so they can be separated
from ordinary search:
chatgpt.com ChatGPT
perplexity.ai Perplexity
copilot.microsoft.com Microsoft Copilot
gemini.google.com Gemini
In Google Analytics 4 you find them under Traffic acquisition by
Session source / medium. It is worth building a separate segment, because for
a long while these sources will be too small to notice in the overall total.
One thing that should not be left unsaid: part of AI traffic looks like
direct traffic. Some clients do not send a referrer header, so those visits
land in the Direct group in GA4. So alongside AI sources it makes sense to watch
direct and branded visits: if those rise while overall traffic stays flat, the
traffic has been redistributed.
2. Search Console: impressions and positions
Impressions and clicks from Google AI Overviews are included in overall Search performance, and there is no separate filter for them. So Search Console does not answer the question "how often was I in an AI answer".
It does answer two others, no less useful:
- which queries have impressions but few clicks — often that means the person got the answer above the results;
- which pages have impressions but return an error or carry
noindex— the one measurement that says "this is costing you already".
3. A citation check against the same list
This measurement is simple and mostly skipped. You write down 10 to 20 questions your customer would genuinely ask, and once a month you ask them, in the same wording, of ChatGPT, Gemini and Perplexity.
You record three things: whether your company was mentioned, which page was cited, and who was mentioned instead of you. The result is a table whose rows can be compared month to month.
The wording has to stay fixed. Change the question and you are measuring something else, and comparison with last month is no longer possible.
4. Technical checks before and after
This is the only part measured immediately, with no waiting. Before the work and after it you record the same numbers:
do answer crawlers get through robots.txt
how many pages carry structured data
how many headings are phrased as questions
how many passages fall within 40-120 words
how many pages return 4xx or carry noindex
These numbers do not depend on a model's mood, which is what makes them the evidence that work was done. A free scan reports them, and the change against the previous scan is shown next to every score.
What cannot be measured
An honest measurement states its own limits:
How many people read the answer without clicking. Nobody sees this except the vendor itself. If your text was cited and the person got their answer, none of your own statistics will show it.
When a model will refresh its data. No vendor publishes a schedule, so the question "how long until it shows" has no precise answer. Technical changes take effect within days, content changes within weeks or longer.
Why one source was cited rather than another. The answer is assembled anew each time, so there is no explanation available that could be verified.
How to write the question list
The list has to be about your customer, not about you:
- questions they would ask before buying ("how much does it cost", "which one should I pick", "is it worth it");
- without your company name — mention the name and the model will answer about you, and the measurement shows nothing;
- including a few whose answer is inconvenient for you — those are the ones that reveal what the model treats as important in your market.
Ten questions once a month takes about twenty minutes. That is the whole measurement job.
Related articles
- How AI cites a page: GEO architecture
- GEO myths that are still being sold
- What to do when AI answers take traffic
- Why GEO needs monitoring, not one audit
The full list of checks is on what we check.
Sources
Frequently asked
How do I see traffic from ChatGPT?
In Google Analytics 4, under Traffic acquisition by `Session source / medium`, look for `chatgpt.com`. The same view shows `perplexity.ai` and `copilot.microsoft.com`. The numbers will be small at first, so a separate segment helps keep them from disappearing into the total.
Does Google Analytics show all AI traffic?
No. Some AI clients do not send a `referrer` header, so those visits land in the Direct group. That is why AI sources are worth watching together with direct and branded visits.
How often should citations be checked?
Once a month. More often is pointless: model data does not refresh daily, and daily checking shows the noise of phrasing and chance rather than a change.
What counts as a result?
Three things in order: the technical checks pass, your company starts appearing in answers to the question list, and traffic from AI sources is no longer zero. The first is measured within a day, the second within weeks, the third last of all.
Check yours
A free scan shows which of the things described here your site already does, and which it does not.