Before anyone rewrites the plan
Your AI visibility numbers didn't move. Now what?
A flat number between two waves feels like a verdict on the plan. Most of the time it isn't - it's a question you haven't answered yet.
28 Labs · August 2026
Flat AI visibility numbers usually aren't proof the strategy failed. Before you touch the plan, run six checks in order: was the measurement itself comparable, did an engine ship a model update between waves, is the movement inside normal run-to-run variance, did the fix you shipped actually get crawled and retrieved, did the plan target the slow layer instead of the fast one, and has enough time actually passed. Most panic rewrites happen because nobody ran the first three.
What are the six checks, in order?
Work through these before you change a brief, fire an agency, or tell a client the strategy didn't work. Each one rules out a different reason the number can sit still - or move - for no reason connected to your brand.
1. Comparability
Was this wave measured the same way as the last one? Same questions, same engines, same denominators. Change any of those between two waves and the number in front of you describes the instrument, not your brand.
2. Model change
Did the engine ship a model or feature update between waves? OpenAI pushed a GPT-5.6 update to ChatGPT on August 6, 2026. Google made Gemini 3.5 Flash the default model inside AI Mode in May 2026. Neither release had anything to do with you, and either one can move your numbers with zero change on your side.
3. Noise floor
AI answers aren't deterministic. Researchers testing an LLM at temperature zero ran the identical prompt 1,000 times and got 80 different completions - an 8% variance rate before anyone changed a single thing. A small move inside that range isn't signal, in either direction.
4. Did the fix actually land
Is the new page live, crawlable, indexed, and actually being retrieved - not just published? A page sitting in a CMS the crawler hasn't visited yet is a fix that hasn't happened. Confirm the page is in the index before asking why an answer doesn't mention it.
5. The wrong layer
Retrieval visibility - what an engine can find and cite right now - can move inside weeks. What a model "knows" from training moves over months, indirectly, as the broader web changes. If the plan targeted the slow layer, a flat number on a four-week check is expected, not failure.
6. Time
Web content typically needs weeks to get recrawled, re-indexed, and re-retrieved before it shows up in an answer at all. Google itself says a page can take anywhere from a few hours to several weeks to get revisited, depending on how authoritative and frequently updated the site is. AI retrieval sits downstream of that same crawl - it can't cite what it hasn't read yet.
Why do flat numbers happen when nothing changed on your end?
Checks one through three exist because a number can move, or refuse to move, without your brand doing anything at all. That's not a caveat we add to soften bad news - it's how the instrument actually behaves.
Comparability breaks more often than people expect. If your first-party citation dashboard adds a new metric or expands its surface coverage between two waves, like Microsoft did when it added Citation Share and Compare to its Bing Webmaster Tools AI Performance report in June 2026, comparing before and after tells you the report changed, not that Copilot changed its mind about you.
3
Only then, question the plan
Skip step one or two and you'll rewrite a strategy to fix a measurement problem. That costs weeks and teaches you nothing.
How long should a fix actually take to show up?
Longer than most briefs assume, and it depends which layer the fix targets. Retrieval - what an engine finds and cites today - is the layer that moves fastest, and it still needs weeks: the page has to be crawled, indexed, and then win the engine's internal retrieval before it shows up in an answer. Training memory, what a model "knows" without looking anything up, moves over months and moves indirectly, as a function of what gets said about you across the open web. We cover that split in detail in what AI visibility tools can and can't measure, because it's the single most common reason a client calls a plan broken when it's actually just early.
A four-week check on a retrieval fix is a reasonable first look. A four-week check on anything meant to shift what a model has learned is checking the wrong clock.
Is a flat number ever actually a real problem?
Yes - once the first five checks come back clean and enough time has genuinely passed. That's the whole point of running checks in order: they exist to tell you when a flat number has stopped being "too early" and started being a real signal.
Not a signal yet
- Different question set or engine mix between waves
- A model or feature update landed inside the window
- Movement sits inside normal run-to-run variance
- The published fix isn't confirmed crawled or indexed
- Under roughly two months since the change shipped
Worth rewriting the plan
- Same instrument, same engines, two clean waves apart
- No model or feature update in the measurement window
- Movement exceeds the noise floor on repeated asks
- The fix is confirmed live, indexed, and retrieved
- Roughly two months elapsed and the number still hasn't moved
The honest readA flat number is information, not embarrassment. It tells you where to look next - the instrument, the calendar, or the plan itself. Most panic rewrites we've seen happen because nobody looked at the first two before reaching for the third.
What we do about it
We run the same question set, on the same engines, wave over wave, so comparability is never the question mark. We log model and feature updates against the run calendar before we read movement as signal, and we treat anything inside normal re-ask variance as noise rather than a story. Every count we publish re-derives to specific questions and specific answers, so when a number moves, or refuses to, you can see exactly why - and whether the plan needs to change or just needs more time.