Suppose you checked last quarter: ChatGPT knew your business, described your services correctly, even cited your website. Great: you passed. Except AI visibility isn't a test you pass. It's a reading on an instrument, and the needle moves whether you're watching or not.
Why answers change under your feet
Every layer of an AI answer is in motion. The models themselves are retrained and replaced; a new version may know more about you, less, or different things. Retrieval changes weekly: engines that search the web before answering re-rank their sources constantly, and the page that anchored last month's answer may not be fetched this month. Your own site changes: a redesign, a moved page, an expired listing can quietly remove the thing an engine was leaning on. And the world changes: new businesses, new reviews, new content all recalibrate what an assistant considers worth saying.
None of these shifts announce themselves. There is no notification when an assistant stops confirming a service you offer. The only way to notice is to ask again: the same questions, the same way, on a schedule.
What a weekly run measures
AnvixaAI's visibility runs re-ask the full battery weekly (every engine, every prompt) and reduce each run to numbers you can chart:
- Knowledge: how much of your prompt set the engines answer correctly and confidently.
- Citations: whether your site is the source engines lean on, or whether they're citing directories and third parties.
- Confidence: hedged answers and misfires, caught even when the surface reads friendly.
The portal keeps the last twelve runs, and twelve points make a trend. A trend is what turns anecdotes into decisions.
The chart you actually want: publish structured data in week 4, watch the engines pick it up over the following weeks.
Weekly, not daily, on purpose
Why not check every hour? Because the signal doesn't live there. Engine answers have day-to-day jitter (sampling, load balancing, retrieval variance) that means nothing. The meaningful movements (a model release, your schema getting absorbed, a source dropping out) unfold over weeks. A weekly cadence samples exactly the frequency where truth lives, and keeps the briefing readable instead of noisy. It's also why each run stores its full evidence: when a number moves, you can open the week and read the actual answers that moved it.
The payoff compounds: after you've invested in machine readability, the trend is your receipt. Structured data ships in week 4; knowledge climbs through week 8; citations flip from third-party directories to your own domain. No screenshot proves that. Twelve weeks of instrument readings do.
- Models, retrieval, your site and the world all shift: answers drift silently
- Fixed prompts on a fixed cadence are the only honest comparison over time
- Weekly sampling matches the frequency real changes happen at. Daily is noise
- Trends turn your visibility work into measurable, attributable results