AI visibility is a signal, not the asset

Rand Fishkin has been asking an uncomfortable question about AI visibility: how much confidence should marketers place in metrics built on systems that give different answers, expose only part of user behaviour and offer very weak causal attribution? I think the question is right. I am less interested in declaring GEO measurement useless than in deciding what should sit underneath it.

Rand Fishkin posted a useful provocation about AI measurement this week.

His argument, in simplified form, is that the foundations are shakier than the dashboards often make them look.

Ask an AI system the same recommendation question repeatedly and the answer changes. Prompt-volume estimates are incomplete. Personalisation and conversation context can alter outputs. A citation does not give you a clean audit trail showing exactly how much that source caused the final answer. And even if a brand appears more often, connecting that visibility to an incremental sale is difficult.

His proposed response is essentially: invest in mentions, look for incremental lift and be very suspicious of the rest. I agree with most of the diagnosis, but I land somewhere slightly different on the conclusion.

I do not think the answer is to stop looking at AI visibility. I think it is to demote it.

AI visibility is useful as a signal. It is not the asset. The asset is the evidence underneath it.

The problem is not that AI visibility tells us nothing

Fishkin’s own earlier research with Gumshoe is important here because it stops the argument becoming too simple.

In a study published in January 2026, 600 volunteers ran 12 recommendation prompts across ChatGPT, Claude and Google AI a combined 2,961 times. The same brand list rarely appeared twice and the same list in the same order was rarer still.

That sounds devastating for AI tracking.

But the study reached a more nuanced conclusion. Individual rankings were extremely unstable, yet the frequency with which particular brands appeared across many runs could still contain useful information. Fishkin wrote that visibility percentage across dozens or hundreds of prompts, run repeatedly, looked like a reasonable proxy for how strongly a brand sat inside an AI system’s consideration set.

That distinction matters.

The weak claim is:

“We rank number three in ChatGPT.”

The stronger claim is:

“Across this defined set of questions and repeated observations, we appear more or less frequently than before.”

The second is still imperfect. It depends on the prompts selected, the models tested, the timing, the users or APIs involved and the way the measurement provider constructs the sample.

But imperfect is not the same as useless.

The real danger is false precision.

A probabilistic observation is being presented as if it were a stable search ranking, and a visibility movement is sometimes discussed as if it were a demonstrated commercial effect.

Those are much larger claims than the data can support.

The output is the least stable part of the system

This has been bothering me because a lot of GEO discussion begins at the end. We look at the answer: did the brand appear, which competitor appeared first, which URLs were cited, was the sentiment positive, what percentage visibility did a platform calculate?

Those are all observable things. They can be useful. But they are downstream of almost everything the organisation actually controls.

An organisation cannot determine exactly what one person’s ChatGPT conversation will generate next Tuesday after thirty previous messages, a model update and a live web search. It can influence the information environment that system has to work with: producing original research, capturing practitioner experience, documenting customer outcomes, making product claims clearer and better supported, putting named expertise into the public domain, maintaining important pages, earning independent corroboration and making evidence easy to retrieve, understand and attribute.

Those things are considerably more durable than a single generated answer. That is why I increasingly think the useful hierarchy is:

Evidence → activation → AI observation → business outcomes

rather than:

AI visibility score → content production → claimed ROI

The order changes the behaviour.

I am trying to move GEO upstream

I have been describing part of my recent working method as becoming a GEO journalist. The name is slightly awkward, but the behaviour is useful.

Instead of beginning with “what article should we produce to rank in AI?”, I am trying to begin with a different set of questions: what are people actually trying to understand, where is the current answer weak, what does the organisation already know, what evidence is trapped somewhere else, what would we need to learn before we could answer properly, who should we ask and what would make the answer more defensible?

That has changed the way I approach fairly ordinary editorial work.

At an accounting industry event, for example, I was less interested in producing another event recap than in capturing practitioner evidence about the operational reality behind Making Tax Digital: where additional work appears, how firms handle clients with multiple income sources and what changes when quarterly reporting becomes part of the normal workflow.

The event was valuable because it produced evidence that could strengthen future answers.

The content came afterwards.

I have been applying the same logic to freelancer content. Rather than treating an advisory board as somewhere to collect colourful quotations after a draft exists, I can take defined evidence gaps into the conversation. If an article is about chasing an overdue invoice from a friend, I can ask what language people actually use, what makes the conversation difficult and what they wish they had agreed before starting the work.

The answer may change the article.

That is the point.

And in customer storytelling, I am experimenting with recording not only the finished story but also the proof points, evidence gaps, useful quotes, themes and questions we wish we had asked. A customer story then becomes reusable organisational evidence instead of a page that disappears into an archive after publication.

None of this guarantees that an AI system will cite the resulting work. That would be the wrong promise. What it can do is create a stronger information environment.

This is where the Evidence Engine matters

I have been building an Editorial Intelligence model I call the Evidence Engine.

The core idea is simple: most organisations already produce large amounts of knowledge through research, customer interviews, events, support, sales, product usage and subject-matter expertise. What they often lack is a system for connecting that material and making it reusable.

Recently I added another layer to the model. When an important question or claim gap appears, do not begin by choosing the format. Choose the evidence.

The working sequence is:

Question or claim gap → evidence type → acquisition → verification → synthesis → activation → reuse

The evidence might be proprietary quantitative research, practitioner reporting, a customer story, expert interpretation or external authoritative data. Often the strongest answer combines several. This is much closer to how I want to think about GEO.

The objective is not to manufacture pages around synthetic prompts. It is to use the answer environment as one source of commissioning intelligence.

A weak AI answer can tell you that the information environment may be weak. A missing brand can suggest that representation is weak. A competitor’s repeated presence can tell you that its evidence, authority, distribution or category association deserves investigation, while a recurring citation can point towards a source class the model currently trusts.

Those are clues, not causal proof.

I still want the measurements

This is the part where I differ from the most aggressive version of the anti-GEO argument. I still want AI visibility data: repeated mention rates across a meaningful set of questions, recurring sources, descriptions of the organisation, whether important facts survive synthesis accurately and whether new evidence starts appearing in the answer environment.

I also want to compare those movements with branded search, direct traffic, referrals, conversion and qualitative buyer feedback where that data exists.

Fishkin himself highlighted Similarweb research earlier this year suggesting that people exposed to brand recommendations in AI were more likely to make subsequent direct visits or branded searches in the three consumer categories studied. He also pointed out the unanswered questions around causality, smaller brands and other channels.

That is roughly the level of confidence I want: an interesting signal worth tracking, not a licence to pretend that the dashboard has reconstructed the sale. The measurement system should preserve those distinctions.

An AI visibility baseline is a sample of representation under defined conditions, not a census of what every buyer saw. A citation is evidence that a source appeared in a generated response, not proof that the source single-handedly caused the recommendation. A rise in direct traffic after increased AI mentions is a pattern worth investigating, not automatically incremental revenue attributable to GEO.

This sounds obvious written out. Dashboards have a habit of making obvious caveats disappear.

The audit should produce questions, not just scores

This also changes how I think about the AI Visibility Audit I have been developing. The most useful output should not be a big number saying your organisation is 42% visible. It should be a diagnosis.

Which buyer questions matter? How are you represented across a controlled sample of them? Where is the description inaccurate, generic or inconsistent? Which evidence appears to support the current answer, what important proof is absent, what does the organisation already possess that has not been routed into the public information environment, what genuinely needs new research or reporting, and what should be retested later?

The visibility measurement becomes the beginning of the work rather than the product. That is why I have also been developing GEO-informed thought leadership: the basic proposition is that GEO should improve commissioning decisions before it improves finished pages.

If an AI answer reveals that every company in a category is repeating the same unsupported claim, the interesting opportunity is not to repeat it more efficiently but to investigate the claim. Maybe the organisation has customer data that can answer it, an expert can explain why the conventional answer is incomplete, practitioners disagree with the accepted wisdom, or nobody actually knows.

Any of those discoveries is more valuable than another optimised paragraph.

There is a bigger trap here

AI has made measurement theatre unusually tempting because the interface feels observable. You can type a prompt, see the answer, count the brands, record the citations and repeat the process at scale. The numbers arrive quickly.

That makes the output feel more measurable than things such as reputation, trust, expertise, customer knowledge or narrative strength. But ease of measurement is not the same as strategic importance.

The same problem has existed in marketing for years. Clicks were easier to count than the effect of a respected analyst mentioning you. Downloads were easier to attribute than the value of a customer repeatedly hearing the same credible story from different sources.

AI does not remove that problem; it may intensify it. We now have a new class of highly legible intermediate metrics sitting between information and behaviour. That is useful, but it is also dangerous if the intermediate metric becomes the objective.

The controllable layer is the interesting one

I keep coming back to a fairly simple split: there are things I can control or deliberately improve, things I can observe, and things I can sometimes infer with appropriate caution.

The first category includes original evidence, customer proof, expertise, clarity, provenance, structure, accessibility, distribution and consistency. The second includes mentions, citation patterns, descriptions, sentiment, referral traffic, branded search and direct visits.

The third is where caution matters most: whether representation appears to be strengthening, whether particular evidence is travelling, whether buyer awareness may be changing and whether a campaign or body of work appears to be contributing to commercial outcomes.

The mistake is collapsing all three categories into one dashboard. Editorial Intelligence is useful to me partly because it keeps the first category visible: it asks what the organisation can learn, capture, connect and improve regardless of whether a model happens to produce the same answer tomorrow.

That is a much more durable place to invest.

AI visibility is still real

None of this is an argument that AI discovery does not matter. People are using AI systems to research problems, compare options, learn categories and form shortlists, and organisations should care about how they are represented there. But representation is downstream of reputation, evidence and information.

So I want to treat AI visibility in the same way I would treat a good journalistic lead. It tells me where to look, gives me a hypothesis, may reveal a gap and may show that something has changed. Then the work begins.

The long-term advantage is unlikely to come from knowing the perfect prompt to track. It will come from becoming the organisation with the strongest, clearest and most reusable body of evidence around the questions that matter. A dashboard can tell you whether that evidence seems to be travelling; it cannot build the evidence for you.

That is why I think Rand Fishkin’s challenge is useful. Not because it kills GEO, but because it forces us to decide what GEO is actually for. For me, the answer is becoming clearer.

AI visibility is a signal. Evidence is the asset.

---

This piece responds to Rand Fishkin’s LinkedIn post and his earlier SparkToro/Gumshoe research on AI recommendation consistency. His June 2026 discussion of Similarweb research on downstream effects of AI brand mentions is also useful context.