Write down what is true
Each fact gets a source, an owner, a start date and an expiry.
Truth registryAnswer integrity
Quality control for what AI says about your company.
Your buyers stopped opening ten links. They ask an assistant, get one answer, and act on it. When that answer is out of date, you lose the deal before anyone reaches your site.
We find the wrong answers across ChatGPT, Claude, Gemini and Perplexity, fix the page they came from, then ask again to check the answer changed.
- We ask
- OpenAI · Anthropic · Google · Perplexity
- Every number shows
- the rate, the error bar, the run count
- Under 5 runs
- no number at all
Slack's free plan works well for a small team. It keeps your 10,000 most recent messages, so nothing is lost while you stay under that. Paid plans add unlimited history and Slack Connect.
| top10teamchat.example.com/best-slack-alternatives | does not support it |
| slack.com/pricing | supports it |
Your approved fact says otherwise
Contradicted
The Slack free plan keeps 90 days of message history. The 10,000-message limit ended on 1 September 2022.
in force 2022-09-01 → current · sensitivity material · approved by pricing-ops
9%95% CI 5%–15%n=116
Critical
Why counting mentions misses this
The answer above is sourced, and wrong.
The brand is named, the tone is positive, and one of the two citations is Slack's own pricing page. Every share-of-voice tool scores that as a win. The limit it quotes ended in 2022.
Check it yourself: ask any assistant what Slack's free plan keeps, then open slack.com/pricing.
Slack's free plan works well for a small team. It keeps your 10,000 most recent messages. Paid plans add unlimited history and Slack Connect.
01
One answer, no second opinion
Search handed your buyer ten links to compare. An assistant hands them one answer. If they act on the wrong one they never reach your site, so nothing in your analytics records it.
02
It used to be true
Slack really did keep 10,000 messages until September 2022, and thousands of comparison pages still say so. Every fact you give us carries a start date and an expiry, because an answer can be correctly sourced and still wrong.
03
The fix is upstream
A stale comparison page carried the old limit and the pricing page did not outrank it. Publish a dated correction, fix the doc, ask again, and see whether the answer moved.
How it works
Five steps, run on a schedule.
Mention counts are an input here, never the deliverable. The loop turns a wrong answer into a corrected one, then into evidence that the fix worked.
Ask what buyers ask
Real buyer questions, asked repeatedly across four assistants. We save the exact setup: model, version, grounding, country.
ObservatoryCheck every claim
Every statement is checked against your facts. Every citation is checked for whether it backs the claim.
Answer deskFix the record
A fixed list of moves: publish a dated correction, fix the doc, add the comparison page. Each carries its evidence.
ActionsProve it moved
Ask again, compare against an untouched baseline, report the difference with a p-value and what we cannot rule out.
ExperimentsSix refusals, enforced in code
Six things this product will not do.
Each one is a failing test. If a release ships a blended score or a bare percentage, the build goes red before you see it.
One score you can screenshot
Ask about you by name and you get mentioned. Ask for a vendor recommendation and you might not. Averaging the two makes a flattering number that means nothing. We keep them apart.
assertNoBlending() throws · tests/unit/intent.test.ts
A percentage with no sample size
Every rate ships with its error bar and its run count. Under five runs you get "insufficient data".
domain/stats.ts · tests/unit/stats.test.ts
An alert because a number wobbled
A change is reported only if it clears a significance test, moves at least ten points, and survives a correction for everything else tested that round.
two-proportion z-test, BH at q=0.1 · services/dashboard.ts
A made-up impact estimate
A fix gets a predicted range only if your workspace already holds comparable experiments. Otherwise it ships as an experiment.
deriveExpectedRange() · tests/unit/priority.test.ts
Reviews and posts we manufacture
There is no connector for posting anywhere, and none for generating reviews. The one review action asks your real customers.
ACTION_TYPES closed enum · tests/unit/product-copy.test.ts
A promise to make the models obey
Nobody outside a lab decides what a model says. We measure it, fix what it reads, and test whether the answers moved. A lint fails the build if this page says otherwise.
banned-claims lint over src/ · tests/unit/product-copy.test.ts
Measurement design
The rules behind every number.
These settings decide what the console shows you. The same page ships inside the product. If a line on it stops being true, that is a bug.
| Fewest runs | 5 per question cluster per window before any rate is shown |
|---|---|
| Most runs | 20, spent where demand, value, volatility and defect risk are highest |
| Error bar | 95% Wilson score interval, still correct at 0 hits and at all hits |
| Change detection | two-proportion z-test, p < 0.05, at least a 10-point move |
| Multiple testing | Benjamini-Hochberg at q = 0.1 |
| Under the floor | no number, just "insufficient data" |
| Saved with each run | provider, model, version, access mode, grounding, search mode, geo, language, personalisation, system config hash, temperature, seed |
| Assistants covered | OpenAI, Anthropic, Google, Perplexity |
Fifty clusters across four assistants for a month costs $400 to $1,000 in model and search spend before anything else. That is why this is not a $49 tool. At $49 nobody can ask often enough to know whether the answer is right.
Pricing
Priced by how many buyer questions we watch.
We load your facts, ask your highest-intent questions, and hand back every wrong answer we can evidence. The report is yours either way.
Start the audit50 question clusters across four assistants, sampled weekly. Alerts have to clear a significance test.
Talk it through100 clusters, sampled daily, plus the full fact registry, the action list and the experiment ledger.
Talk it throughMultiple brands or clients in one place, CRM handoff, approval trails and full export.
Talk it throughStart with the free audit
See what the assistants are telling your buyers this week.
The audit is the whole product, run once by hand on your domain. Most teams have never read a transcript of what the assistants say about them.
- We load your facts from your docs. You approve each one before we ask anything.
- We ask your highest-intent questions across every assistant this deployment has a key for.
- You get every wrong answer we can evidence, with transcripts, setups and citations.
- If we find nothing worth fixing, we say so and you owe us nothing.