Searchable's Chris Donnelly: a near-100% AI score missed the questions that mattered
Searchable's Chris Donnelly explains why brand-heavy prompt lists can make AI visibility scores look nearly perfect while missing new buyers.
A manufacturer was getting an almost perfect mark for how often AI mentioned it. The company's prompt tracker showed visibility close to 100%. There was just one problem: some of those questions still concerned a division it had sold earlier in the year.
Remove the old and current brand names, ask the category questions a potential customer would ask, and the result looked entirely different. The company's visibility on that new set was almost zero. It was a measurement problem hiding in plain sight. The tool had been looking for the company where it was already easy to find.
That is the problem Searchable, a platform for tracking how brands appear in AI search, is targeting with its Prompt Universe feature. ContentGrip spoke with Chris Donnelly, the company's co-founder and CEO, about why an impressive visibility score can say so little about the people who have not chosen a brand yet.
Table of contents
- The question map comes before the score
- How a manufacturer went from almost 100% to almost zero
- The 500-prompt list that missed a fashion market
- Why a SkyMiles question is a bad test for Delta
- Why one sportswear brand chose 1,200 questions
- What to ask before trusting an AI visibility score
The question map comes before the score
Searchable introduced Prompt Universe on September 30, 2026, to map a wider set of possible buyer questions before selecting which ones to monitor. Instead of starting with a marketer's chosen list of brand terms, it models questions across the company's products, audiences, rivals, and purchase situations. It then trims that map to a manageable sample.
The larger number is not a count of real ChatGPT conversations or of searches people actually made. It is a modeled universe of questions that could arise. The point is to discover what the original list failed to ask before committing to a regular tracking plan.
The source material for each map includes the company's own site and language found in buyer forums and communities. Chris also described looking at searches AI assistants run while answering and using anonymized information about how people phrase questions. The team then removes combinations unlikely to make sense for the company or its customers.
Uber offers a concrete illustration. In a sample analysis of the ride-hailing company, Chris said Searchable considered the concerns riders raise about "surge pricing, cancellation fees, or airport pickups". Those situations are more informative than generating hundreds of interchangeable variations on the word Uber. This was a sample analysis of a named brand, not a claim that Uber was a Searchable customer.
As co-founder and CEO, Chris's point is that the choice of questions shapes the score before the first AI answer is collected.
How a manufacturer went from almost 100% to almost zero
The unnamed manufacturing client had been tracking questions about its current brand and products, plus a business division it no longer owned. Those prompts made its visibility look close to 100%, Chris told ContentGrip.
Searchable then tested a different question set, removing brand and product names and asking about the categories and buying situations that would matter to a new customer. Visibility fell to almost zero. This was not evidence that the brand suddenly lost ground in AI search: the team had changed what it measured.
Chris said a branded-heavy list "mostly measures whether AI knows you exist, instead of whether it recommends you to a stranger." The client was advised to keep a smaller watch on those familiar branded questions and pay more attention to the categories in which the business was missing from AI recommendations.
That distinction makes the number useful. Branded recognition can matter, but a buyer who already knows your name presents a different problem from someone asking AI which manufacturer to consider.
The 500-prompt list that missed a fashion market
A second client, an unnamed multibillion-dollar fashion company, arrived with roughly 500 tracked prompts. The list favored its own name, hero products, and predictable questions such as "best handbag". Those are useful questions, but they describe only part of how people shop.
When Searchable mapped the wider market, Chris said it found more than 20,000 realistic possible questions. The blind spots were recognizable shopping moments: choosing a gift, buying for a season, comparing specific competitors, checking authenticity, or worrying about what happens after purchase.
On the unbranded questions the team tested, a competitor came out ahead on more than half, according to Chris. The fashion company rebuilt monitoring around buying situations and topics instead of one flat prompt list. Its content plan also shifted toward comparison and decision-stage questions that had previously gone unmeasured.
These figures are Searchable's account of an anonymized client engagement, not independently audited outcome data. What they show is the practical risk of sampling only questions a brand team expects to win.
Why a SkyMiles question is a bad test for Delta
For Delta Air Lines, Searchable modeled more than 115,000 possible combinations. After filtering for realistic questions, its published example retained more than 33,000 for consideration. Those were candidate questions, not a mandate to pay for tracking every one.
The team kept travel choices such as flights between New York and Paris, plus questions about basic-economy fares or choosing an airline for a family trip. Each could expose an actual comparison between airlines.
But a question about the cost of joining Delta's SkyMiles program was a poor unbranded test. SkyMiles already points to Delta. Chris said the more useful version was "What does it cost to join a frequent flyer program?" That leaves the assistant to decide which program or airline to mention.
This is where the prompt filter does substantive work. Simply removing a company's name from a spreadsheet does not always remove clues that make its appearance nearly inevitable.
Why one sportswear brand chose 1,200 questions
More questions are not automatically worth monitoring. Chris described a sportswear company whose universe held about 6,900 potential prompts. Tracking even half would have stretched its budget and marketing priorities.
Instead, the company chose one commercially important topic and monitored roughly 1,200 questions in depth. They were unbranded, tied to comparison or buying decisions, relevant to products sold in the company's main markets, and included situations where competitors might appear.
Chris's argument is that a focused sample works when marketers know the wider market it represents. It is less helpful when the sample is small simply because somebody guessed the first few hundred questions to type into a dashboard.
There is another reason to be cautious about a precise-looking score. AI answers can change across runs and conditions. A measurement guide from Supermetrics recommends repeated observations and a consistent prompt set rather than treating one answer as a stable ranking. Expanding the set may improve what a team learns, but a new set should not be compared with the old one as though the underlying test were unchanged.
What to ask before trusting an AI visibility score
The first audit is straightforward: look behind the percentage. How many tracked prompts already contain your company name, products, or proprietary terms? How many resemble the questions of people who have never heard of you?
Then ask the people closest to customers. Sales, support, and community teams can identify the awkward comparisons, price concerns, use cases, and post-purchase worries that rarely appear in a polished marketing keyword list. Group them by topic and commercial importance before deciding how many to track.
Finally, keep branded recognition and unbranded recommendation performance separate. They answer different questions. The manufacturing client's almost-perfect score was a real result for its old sample, but it concealed where the business was absent from the decisions a new buyer might make.
The lesson from Chris's cases is not that every brand needs 20,000 prompts. It is that a score should not be trusted until the marketer can explain why those particular questions were chosen, who might ask them, and which business decision the answers can change.
