Every AI visibility platform on the market reports a percentage: how often models mention your brand. It is a clean metric and it presents well to a board. On its own, however, it is among the least useful things you can know about your position in AI search. I say that as someone who runs the commercial side of a company that sells exactly this kind of measurement.
I recently hosted a webinar with Dean Harris, founder of The Navigators, a brand research and measurement consultancy based in Sydney with an office in New Zealand. Dean has been deploying our platform with clients across Australia and New Zealand for six months. My colleague Jorge Soto moderated. What follows draws on that conversation and on what our own platform data now shows.
What 175,000 personas reveal about the average
We have seen approximately 25,000 projects created on the platform and rerun well over 50,000 times, generating more than 175,000 buyer personas. That is a substantial evidence base.
The most consequential finding concerns variance. Visibility differs by 21 to 40 percentage points between personas for the same brand. In one instance, a fitness apparel company, a key buyer persona in their category appeared in 16% of answers while a second persona appeared in 80%.
The underlying mechanism is familiar to anyone who uses these tools. Whichever model you rely on adapts its responses based on what it knows about your work, your interests and your circumstances. It is attempting to give you the correct answer, not the average one. What we did not know until we had the data was the magnitude of the resulting spread.
Most GEO trackers do not model the buyer at all. They execute a list of prompts, aggregate the results and report the mean. That approach discards the detail, and with the detail goes the fidelity, the opportunity and the specific tactics that would tell you what to address.
When AI describes a customer you do not recognise
Our platform surfaces the buyer personas that models infer for a category before it measures anything. In early demonstrations, when personas returned that did not match who a prospect believed they were selling to, my assumption was that the platform had erred.
It had not. The platform surfaces what AI understands, and AI is a representation of the world, so what you are seeing is the conclusion the market has reached about your business. If models believe you sell to one set of buyers while you are confident you sell to another, that is not a data quality issue. It is a strategic alignment issue. If AI does not understand you, or does not understand who you sell to, the visibility percentage is the smaller of your problems.
Dean treats this as foundational. Before running anything for a client he establishes the category definition, then places the inferred personas alongside the client’s existing segmentation and works through it until both parties agree. He does this in every engagement, for two reasons. The measurement is only as reliable as the definition of the buyer. And clients who have contributed to the personas take the findings considerably more seriously afterwards. In his words, it gives them skin in the game.
He has also found the exercise surfaces audiences the marketing team had not been considering, which has caused several of his clients to broaden how they define their market.
There are two reasons a model will not recommend you
If AI is not recommending your brand, one of two conditions applies, and they require opposite responses.
Either the model lacks what it needs from you, meaning your content does not address the question or is not structured in a way that supports retrieval, or the model holds sufficient information and has determined you are not the appropriate choice for this buyer. The first is a content problem, addressed by publishing. The second is a perception problem, and additional publishing will not resolve it.
Conventional GEO tracking cannot distinguish between the two. Those trackers request recommendations, and models name only brands they regard positively; they will not recommend a brand about which they hold a negative perception. Consequently, everything being quietly passed over registers as a low number with no explanation attached.
This is why we built a sentiment audit that generates a different class of prompt. Should I choose this brand or that one. For this particular use case, which of these is appropriate. Comparative questions, designed to elicit positive and negative attachment rather than a shortlist. The results indicate whether models lack information about you or are actively directing buyers elsewhere.
Our content audit addresses the same question from the opposite direction: do you hold content that answers what buyers are asking, and if so, is it being retrieved and cited?
The visibility audit answers what. These answer why.
AI search moves faster than search ever did
Results in AI search shift considerably faster than in traditional search, and this cuts in both directions. Challenger brands can take share more quickly than history would suggest. Category leaders can lose it just as quickly.
We tested the speed directly. Dean had identified YouTube as an increasingly significant source, particularly for Google’s AI Overviews, which is consistent with what we observe in citation data. We therefore produced content for Gumshoe and published it to YouTube, constructed so that the title and body carried semantic similarity to the prompts our personas actually ask. It was cited within 18 hours. View volume was immaterial. Where content addresses the question a model is attempting to answer, it gets used.
The work changes depending on where you already stand
One of the more practical observations Dean offered is that the required work differs substantially by starting position.
For brands with low visibility, the constraints are typically internal. His example is a large national not-for-profit childcare provider whose site did not support AI ingestion, lacked the schema that helps models interpret content structure, and carried content misaligned with what parents are seeking. Competitors performing well had pushed content down to the individual centre level, with Google reviews and testimonials from parents at that centre published on that centre’s page, because that is where the research is grounded: a provider nearby that can be trusted. His client held that material on corporate pages instead.
For the middle tier, approximately 20 to 30% visibility, where most of his clients sit, sites are generally ingestible and the gap is off-platform. The conversations shaping the category occur in third-party sources the brand does not own. The work becomes citation analysis and earned media: establishing which sources and which creators are gaining traction with models in that category. Dean is applying this with a PR agency whose global pharmaceutical client is developing a new cancer treatment, where the audiences are narrow and specific, including cancer researchers and biotechnology investment analysts.
For a category leader, the work is optimisation and defence. His medical indemnity insurance client recorded approximately 92% visibility across five models, which suggests little remaining opportunity. The visibility was not evenly distributed. Certain physician personas encountered them far less frequently, and Google’s AI Overviews lagged their presence in other models. The work became closing those specific gaps, supported by competitive intelligence on which rivals the models were rewarding and why.
This is not SEO with a different letter on the front
Dean articulated it more precisely than I have: this is not SEO with a different letter at the front of the acronym, it is human and brand insight in an AI-mediated world. Models are not a field into which users enter keywords to receive links. They build context and understanding, and they calibrate their responses to their assessment of the person on the other side of the screen.
The discipline of marketing has not changed. It was never about producing content or moving product; it is about satisfying customer needs. Dean’s framing is that you become more visible and more authoritative to a model by increasing the overlap between what your content says and what your customers require. The greater that intersection, the better your prospects of forming part of the answer, and of being recommended ahead of the alternatives.
Common questions
What is GEO, and how does it differ from AEO?
The terms are largely interchangeable. GEO, generative engine optimisation, concerns how a brand appears within answers produced by AI models and AI search experiences. AEO, answer engine optimisation, emphasises being the source of the direct answer. Both describe the same shift away from ranked links towards synthesised responses.
Why measure by persona rather than by prompt?
Because models tailor answers to their assessment of who is asking. Aggregating prompts into a single score erases 21 to 40 point differences between buyer types, and those differences are where the actionable insight resides.
How do you query models that personalise responses?
We are integrated with 11 models through their APIs. The alternative, spawning a headless browser or incognito window and entering a prompt, produces a user with no context, whereas most people using these tools hold an account, are logged in, and accumulate context over time. Working through the API allows us to construct persona context deliberately, load it into an agentic synthetic user, and let that user conduct the conversation. We ask a follow-up on every answer, requesting a rank order of the brands recommended and the reasoning, and the output is reviewable model by model, persona by persona, prompt by prompt.
Can the personas and prompts be edited?
Yes. Everything the platform generates is a starting point. Personas can be edited, deleted or replaced with your own ICPs, topics can be added, and you can write your own prompts. Dean conducts a considerable amount of that tuning on behalf of his clients during onboarding.
If you would like to discuss any of this, or you run an agency and want to hear about our partner programme, I am at jim@gumshoe.ai.
