How to Measure Generative Engine Optimization (GEO)

You measure generative engine optimization (GEO) in three layers: visibility (are you named or cited in AI answers), referral (do those answers send visits) and revenue (do those visits convert). Each layer needs its own metric, measured per assistant, with a baseline taken before you change anything. Never report one layer as evidence for another.

There is plenty of advice about how to get cited by AI assistants, and much less about how to tell whether any of it worked. Search engine optimisation had rank tracking within a couple of years of becoming a job. GEO is several years in, and most teams still cannot answer "did last quarter's work do anything?" This is a framework for answering it, with the failure mode each layer is designed to catch.

For two platforms, partly. For the rest, no.

Google. Search Console's Generative AI performance report, rolled out to all sites on 31 August 2026, shows impressions from AI Overviews and AI Mode, by page, device and country. Google's own guidance on generative AI features says the usual SEO best practices "continue to be relevant" and that there is "no special schema.org markup you need to add".

Microsoft. Bing Webmaster Tools has an AI Performance report covering citations in Copilot, Bing and select partner AI experiences. In June 2026 it added Citation Share: the percentage of citations for a grounding query that point to your site.

ChatGPT, Claude and Perplexity. No reporting surface for site owners. Everything you learn about them comes from sampling their answers from the outside and watching your own site.

So GEO measurement is still largely external. Being clear about which numbers come from a platform, which come from sampling and which come from your own site is what separates a measurement programme from a dashboard of comforting numbers.

Access. An assistant can only cite a page its search crawler has fetched, and AI crawlers read the raw HTML: Vercel and MERJ found that none of the major AI crawlers render JavaScript. If your robots.txt blocks a search crawler such as OAI-SearchBot, or your content only appears after scripts run, no amount of content work will show up in the layers below.

Check this first, and keep checking it. Crawl access breaks quickly and silently, and the effect reaches citations weeks later. Crawls are only visible on your server or edge, never in browser analytics, which is why AI crawler tracking is its own measurement.

The question: when someone asks an assistant a question in your category, are you mentioned?

For Google and Bing surfaces, start with the platform reports above. For ChatGPT, Claude and Perplexity, the only method is synthetic sampling: build a prompt set, run it on a schedule against each assistant, and record what comes back.

The design of the prompt set determines everything, and most teams get it wrong in the same three ways.

Too few prompts. Assistant outputs vary from run to run. A single check measures noise. You need enough prompts, run often enough, that you are estimating a rate rather than observing an event.

Prompts phrased like keywords. "AI attribution tool" is a search query. People ask assistants things like "we can't tell how much revenue comes from ChatGPT, what should we use?" Prompt sets imported from a keyword list measure a behaviour that does not occur.

Only high-intent prompts. Category questions ("what is X", "how do I measure Y") matter as much as purchase questions, because that is where an assistant forms its picture of who the credible players are.

Track three things per prompt, per assistant:

• Mention rate: how often you appear at all.
• Citation rate: how often you appear as a linked source, which is rarer.
• Position: where in the answer you appear, since the first name carries more weight.

Retrieval and citation are also different events. AirOps analysed 548,534 pages retrieved across 15,000 prompts and found that only 15% of the pages ChatGPT retrieved appeared in its final answers. Being read is necessary, not sufficient.

Traceten's AI visibility tracking is in Beta and follows this design: prompts you write, asked on a schedule, with each engine reported on its own.

The question: do citations produce visits?

This is where GEO measurement meets the attribution problem. Many AI-referred visits arrive with no referrer, especially from assistants' mobile apps, so counting only visits whose referrer is an assistant captures a minority. Google Analytics now has an AI Assistant channel, but it only recognises visits whose referrer matches Google's list; what that channel still misses matters here.

What to track:

AI sessions by assistant, with confidence. Not a single number: a count per assistant, named from the referrer, UTM tags and user agent, with a confidence and method on each, so you can tell growth from a detection change. This is what AI traffic tracking in Traceten reports.

Landing page distribution. Which pages assistants send people to. This is the most useful diagnostic in the framework, because it is the only direct link between content and outcome. If your comparison page gets visits and your pricing page does not, that is an instruction.

The unattributed share. Direct traffic landing on deep pages, tracked as a trend. It is a proxy, not a measurement, but its direction is informative.

The question: are these visits worth anything?

Same machinery as any other channel, with two AI-specific complications.

The window has to be long. The AI touch is a discovery touch. A 24-hour or single-session window excludes it and will report that the channel produces nothing.

The model cannot be last touch alone. Last touch erases this channel, because the final touch before purchase is usually a branded search or direct visit. Report an introduction view (revenue where an AI touch appears anywhere in the history) beside a last-touch view. The gap between them is itself the finding.

Setting measurement aside for a moment: the published research on what gets cited is more consistent than the field's noise suggests.

The first two figures come from Aggarwal et al., who tested nine rewriting methods on a benchmark of generative engine queries. Their "top-performing methods, Cite Sources, Quotation Addition, and Statistics Addition, achieved a relative improvement of 30-40%" on their main visibility metric. The effect was largest for lower-ranked sources, which is good news for a smaller site. The third comes from Ahrefs' study of 75,000 brands, which found branded mentions across the web tracked AI Overview visibility much more closely than backlinks did.

Treat all of these as directional: one is a benchmark study and the other is a correlation. But the pattern is coherent: content that is specific, sourced and talked about elsewhere gets cited more than content that is vague and unsourced. The same care goes into writing pages AI answers cite.

Two further findings are worth knowing:

Rankings matter less than they did. Ahrefs found that about 38% of AI Overview citations came from pages ranking in Google's top 10 in early 2026, down from about 76% in July 2025. Ranking still helps, but citations increasingly come from deeper in the results.

Third-party mentions carry weight. Given the mention correlation above, being described by others is a different and stronger signal than describing yourself.

Two rules prevent most self-deception.

Never claim a layer you did not measure. If you measured mention rate, report mention rate. Do not call it traffic, and do not put it on a slide next to a revenue figure without a clear separation.

Establish the baseline before you start. The most common failure in GEO measurement is starting to measure after the work has begun, leaving nothing to compare against. Run the prompt set for a month before changing anything. That month is worth more than the next six of optimisation, because it is the only thing that lets you attribute a change to a cause.