Your AI visibility number was 27. Now it's 13. Something must be wrong with your website — right? That's the reflex, and it's exactly the reflex a lot of vendors are counting on. A falling number is the easiest thing in the world to sell against: it looks like an emergency, and the emergency conveniently requires the work they were going to pitch you anyway.
But a score can drop for two reasons that look identical and mean the opposite. Either your business genuinely became less visible to AI — a real problem worth fixing — or the measurement itself changed: more questions in the test, a stricter rule for what counts, a new AI engine added to the panel. Same line going down. Completely different truth underneath.
A number falling doesn't tell you whether your business got worse or your ruler got stricter. Only checking the record does. And the company selling the fix has every reason not to check.
The two causes, and why they demand opposite responses.
If the drop is real — the engines genuinely stopped naming you — then the fix is to work on the pages and re-measure. Fair enough.
If the drop is an artifact of the measurement — you added five more questions, or you tightened what "named" means, or you started counting a sixth engine — then editing your website fixes nothing, because nothing about your business actually moved. Worse: chasing a phantom drop with page edits corrupts the very instrument that's supposed to tell you the truth. You'd be sanding down a ruler to make a board look longer.
This is why the ruler has to be frozen and disclosed. If the company measuring you can quietly change the test and then point at the drop, the number means nothing — and it's a perfect way to manufacture urgency on demand.
1. Compare on the same frozen ruler. Put the two readings side by side and flag any change to the question set, the scoring rule, or the engine panel. If the ruler moved, the readings aren't comparable — full stop.
2. Separate the pillars. We never publish one blended grade that can hide the truth. Answer-engine visibility (Perplexity, Google AI Overviews — where buyers actually are), model-memory recall (what an engine says with no browsing), and local presence each move on their own. A drop in one is not a drop in all.
3. Only a drop that survives both checks is real. And only a real drop is worth changing a page over.
A real one we caught — and edited nothing.
Here's a live example, on the record. On our founder's own 40-year signage company, Brand 9 Signs, one pillar of the score dropped hard in a single day — from 27 to 13. If we sold panic, this is where we'd have called it an emergency and started rewriting pages.
Instead we checked the record — and the drop landed on the exact day we made our own test harder: we grew the question set from five to thirteen, and we stopped giving credit when an engine only named the business after browsing the web (that's not memory, so it shouldn't count as memory). The business hadn't slipped an inch. On that same day, the channels buyers actually use rose — answer-engine visibility climbed and has kept climbing since. The drop was our ruler getting stricter, not the business getting weaker. So we restated the number honestly and edited zero pages. Every step of that is dated and public.
That's the whole difference. A leaderboard would have used the drop to sell you work. A registry uses it to tell you the truth — even when the truth is "our own measurement moved, and your business is fine." We publish the diagnosis, not just the number.
So before you rebuild anything because a score fell: ask whether the business changed, or the ruler did. If the company that measured you can't answer that with a dated, frozen, side-by-side record — the drop they're pointing at was never really a measurement to begin with.
Are you the answer?
Get a score on a frozen, disclosed ruler — one that tells you the truth when it moves, in either direction, instead of manufacturing an emergency. Public, dated, and reproducible.
Are you the answer? Find out