seomodoI love a dashboard. I also know how easy it is to make one look convincing. Put 73/100 in a large font, add a tasteful gradient, and suddenly it feels like an objective truth. It isn't, unless you can explain exactly what produced the number.
A score is the end of a calculation, not the beginning of the story.
Ask what was tested. Which buyer questions? Which providers and models? Was search enabled? What counted as a mention versus a recommendation? How were citations handled? What happened when a provider failed to answer?
If the answers aren't available, you're buying a number without its ingredients. That doesn't automatically make it useless, but it makes the result difficult to evaluate.
- Prompt coverage: the exact questions and why they matter to buyers.
- Evidence: dated responses, sources, and the difference between mentions and recommendations.
- Scoring: which observations count, what weights apply, and what happens when data is missing.
- Consistency: repeated prompts and an explanation of what can change between runs.
More AI responses do not automatically mean more certainty.
AI responses vary by wording, timing, model, search mode, location and provider. Running a larger set of prompts can offer a broader view, but it does not turn those results into a representative sample of every person's AI search experience.
Seomodo records 188 answers in its complete audit, but it does not score all 188 as independent evidence of market share. Its fixed 12-question unbranded baseline on each of four platforms is the comparable scoring set; additional questions, repeats and branded prompts are recorded for diagnosis. That distinction matters.
The tests use search-enabled APIs. They are not a live feed of everything people see inside consumer versions of ChatGPT, Claude, Gemini or Perplexity. Those products can show different results.
Missing data should not magically become failure.
Suppose a provider times out, a search mode is unavailable, or the expected baseline is incomplete. Counting the missing response as zero can punish a business for a measurement failure.
A responsible report separates 'not observed' from 'not measured.' It also tells you when a comparison is unsuitable because the prompts, model, provider or coverage changed. Measurement coverage is useful context; it is not a statistical confidence interval.
Business impact needs its own column.
A mention does not equal a lead. A lead does not equal a customer. And a modeled revenue scenario does not equal revenue booked.
I want these presented separately: what appeared in the AI answers, what changed in identifiable visits or inquiries, and what the potential value might be under stated assumptions. That's less flashy than one all-powerful score. It's also how you avoid confusing observation with return on investment.
Before you approve a visibility budget, insist on seeing the evidence alongside the score.
The best AI visibility score is the one you can challenge, reproduce and interpret.