Search measurement used to have a centre of gravity. Google held the overwhelming majority of queries, Search Console reported on it directly from the source, and the industry organised itself around that single reference point. Arguments were about interpretation, not about whether the underlying number existed.
That centre is gone, and nothing has replaced it. There are now several major AI answer surfaces, each with different citation behaviour, none with an equivalent of Search Console, and no shared definition of what visibility even means. Semrush's 2026 AI Visibility Index found that 45% of marketing leaders could not accurately measure their brand's presence in AI-generated answers, and only around 9% had tools to track it across platforms.
That gap is not a tooling problem waiting for a vendor to solve. It is structural, and the sensible response is a measurement model that works without a single source of truth rather than a search for one.
The engines behave differently enough to break averages
The most important finding for anyone building a measurement framework is that citation patterns vary enormously by platform.
Semrush's analysis of 126 million US AI search prompts found ChatGPT citing an average of around 15 sources per response, drawing heavily on community and reference platforms including Reddit and Wikipedia. Gemini averaged roughly 3 sources, from a smaller pool that included Wikipedia, Reddit and YouTube.
Sit with that difference for a moment. One engine is assembling answers from a wide field where a strong presence in community discussion can get you included. The other is drawing from a narrow set where being outside the top handful of authoritative sources means invisibility. These are not variations in degree. They reward substantially different work.
The practical consequence is that a blended "AI visibility score" averaged across engines is close to meaningless. A brand can be highly visible in one and absent from another, and the average will describe neither situation. Per-engine measurement is not a refinement — it is the minimum viable approach.
Four metrics, tracked separately
A workable framework separates things that are usually collapsed into one number.
Presence rate, per engine. Of your tracked prompts, what share produce a response mentioning your brand? Track this individually for each platform you care about. Never average them.
Citation rate, per engine. Of the responses mentioning you, how many actually link to your site? This is the metric that determines whether AI visibility produces attributable traffic or only awareness you cannot measure downstream.
Source composition. When the engine answers a query in your category, which domains does it draw from? This is the most actionable thing you can track, because it converts "we are not visible" into a specific list of places to earn presence.
Competitive position within responses. When a competitor appears and you do not, what is the source behind their inclusion? Not their overall visibility score — the specific page or mention that got them there.
Applying reporting discipline familiar from SEO helps here, particularly the habit of reporting on movement and cause rather than on absolute values that no stakeholder can benchmark.
Building a prompt set that means something
Everything downstream depends on the prompts you choose, and this is where most programmes go wrong quietly.
Keyword thinking does not transfer. A buyer does not type "enterprise backup software" into ChatGPT — they describe a situation and ask what to do about it. Your prompt set needs to reflect actual conversational input, which means it should come from your sales team and your support inbox rather than from a keyword tool.
A defensible set covers four types:
- Category-entry prompts. "What are the options for X?" These reveal whether you exist in the model's picture of your market at all.
- Comparison prompts. "Is A or B better for Y?" These show how you are positioned relative to named competitors.
- Constraint prompts. "What should I use if I have [specific limitation]?" These are where mid-sized brands most often win, because the answer requires specificity that generic market leaders do not provide.
- Branded prompts. "What is [your company] good at?" These test whether the model describes you accurately, which is a different problem from whether it mentions you.
Fifty to a hundred prompts is enough for most businesses. Precision in what you ask matters far more than volume, and a hundred well-chosen prompts tracked consistently beats a thousand generated automatically.
Accept variance rather than fighting it
AI responses are not deterministic. Run the same prompt twice and you may get different brands, different sources, and a different ordering. This is genuinely uncomfortable if you are used to rank tracking, and it changes what a measurement actually means.
Three adjustments make the data usable:
Measure frequency, not position. "We appeared in 31 of 50 responses this month" is a stable, meaningful statement. "We were the second recommendation" is a single sample of a distribution.
Use wide comparison windows. Week-over-week movement in this channel is mostly noise. Month-over-month is roughly the shortest interval worth interpreting, and quarterly is where genuine trends become visible.
Annotate everything. Record content launches, PR placements, product changes, and known model updates against the timeline. Without annotations you cannot separate the effect of your work from a vendor changing their retrieval logic, and that distinction is the whole point of measuring.
Where the leverage actually sits
A finding worth building strategy around: analysis of AI citations has repeatedly pointed to earned media and third-party sources carrying far more weight than owned content. One study of 25 million cited links attributed the large majority of AI citations to earned media rather than brand-owned pages.
If that holds for your category — and it is worth verifying with your own source composition data rather than assuming — then the implication is uncomfortable for teams whose entire content investment goes into their own site. The work that moves AI visibility looks more like public relations, community presence, analyst relations, and getting included in the comparison articles and reference resources the models already trust.
That reframes what a GEO programme even is. It is less a content calendar and more a presence strategy, which is why it usually needs to be built into how generative engine optimization fits wider brand and demand marketing rather than run as an isolated SEO subtask.
Reporting upward
Executives will ask for one number. Resisting that request politely is part of the job, because the single number does not exist and inventing one guarantees you will eventually have to explain why it moved for reasons unrelated to your work.
A three-line report works better than a composite score:
- Presence rate per engine, with direction of travel
- The two or three source domains driving competitor inclusion where we are absent
- What we did about the last report's finding, and whether it moved
That structure keeps the conversation on causes rather than on a score, and it survives contact with a vendor changing their methodology.
For teams establishing a programme from scratch, our guide to building an AEO strategy around questions and evidence covers the content side, and the visibility revenue calculator helps size the opportunity before you commit budget to measuring it.
The honest position
Nobody has this fully solved. The tools are young, the engines are changing their behaviour, and the research base is a year old at most.
What separates teams making progress from teams generating dashboards is not tooling sophistication. It is whether the measurement is connected to a decision. If your report cannot name the specific thing you will do differently next month, the measurement is not yet worth its cost.
Originally published on the Mustard Seed blog.
Top comments (0)