Why Two AI Visibility Tools Give You Different Numbers
The same prompt sent to the same engine twice returns different answers — that is how these models work, not a bug in anyone’s tracker. Tools then diverge further on sampling rate, geography, personalization state, and what they count as a mention. Expect disagreement on absolute numbers. Compare trends and relative share instead.
A usable prompt set has 30–50 prompts, zero brand names in the core index, a frozen list, and more decision-stage questions than vanity branded queries.
A vendor that samples 100 times per prompt is estimating a distribution.
Comparing their “mention %” to a once-a-day tracker is a category error.
2. Geography
Exit country changes shopping, language, and which sources get retrieved.
A US datacenter and a DE residential IP are not the same test.
If the vendor will not say where the request leaves, the number is not portable.
3. Personalization and session state
Logged-in ChatGPT with memory is not a measurement instrument.
It is a conversation.
Serious trackers use a clean context.
If one tool is closer to “what your intern saw,” it will flatter or panic you at random.
4. Provider path
API, official live UI, and a scraper are different surfaces.
They see different tools, different browsing, different safety layers.
LLMConquer submits work asynchronously through Cloro (POST /v1/async/task). We do not use Cloro’s synchronous monitor routes for the daily run — those cost extra provider credits and hold concurrency.
Another vendor’s “live ChatGPT” number is a different path.
Named engines on Starter: ChatGPT, Perplexity, Gemini, Google AI Overviews. Pro and Business add Microsoft Copilot.
Cadence: one scheduled pipeline per paid account, 02:00 UTC daily (mention-runner-daily). That is one draw per tracked prompt per included engine, not a confidence interval.
Path: async Cloro task, webhook on completion, then Haiku extraction into mention_results, citations, and competitor appearances.
We do not publish per-cell variance today.
If yesterday you were at 40% and today you are at 28%, that is not automatically a loss.
Look at the week and at who else moved.
Access and rendering still come first. If the bots cannot read you, no tracker can save the content. See the 7-layer diagnostic.
Should you run two tools?
For most teams, no.
You will spend the second invoice arguing about two samples of the same noisy process.
Run two tools when you are in a bake-off with a written end date, or when a number must survive a legal or board challenge and you need an independent method.
Otherwise pick one methodology, freeze the prompt set, and watch relative share.
Frequently asked questions
Which tool is most accurate?
No one has published a study that proves a winner. Accuracy is the wrong word for a sample from a nondeterministic model. Ask how they sample, from where, and what they count as a mention.
Why did my score drop with no site changes?
A single daily draw can miss you. Competitors can ship a comparison page. The model can rotate cited sources. Check a seven-day window and relative share before you rewrite the homepage.
Can I verify my tool’s numbers manually?
Yes, poorly. Run the same frozen prompt logged out, twice, in one region. You will not match the dashboard. You can confirm direction. That is enough to catch a broken integration.
How much fluctuation is normal?
Day-over-day mention percentage is noisy. A one-day dip is not a finding. A four-week slide in rank order versus the same competitors is.
Should I buy two tools to triangulate?
Usually no. You will pay twice to argue about two samples. Use two tools for a time-boxed bake-off or a claim you must defend. For weekly reporting, pick one method and freeze the prompt set.
Do I need a tool at all?
Not until you have a stable prompt set and a real gap. Start with the diagnostic and the prompt-set guide. Buy cadence after that.
Twelve AI visibility platforms compared on a normalized cost: monthly price divided by included prompts × engines × refreshes per week. Prices checked 21 Aug 2026.
Work access, rendering, entity, corpus, competitive, prompt-fit, and freshness in order. Most ChatGPT invisibility is a configuration problem, not a content problem.