Catch AI answers that get your facts wrong
Keep a library of your prices, dates and other key facts. Measure checks every answer it collects for your tracked prompts and shows which engine got what wrong, next to the truth.
Product preview, illustrative data
What it does
A fact library built to be checked
Store up to 60 facts per project as numbers, percentages, dates, yes/no answers, URLs or short phrases. Each has anchor words and a critical or normal flag, and numbers also take a unit and tolerance.
Deterministic, not another model's opinion
No model grades the answers. Measure reads the value from the first sentence about you that mentions each fact and compares it by type: numbers with unit scaling and tolerance, dates, yes/no with negation, URLs.
The wrong sentence next to the truth
Each contradicted fact shows which engines got it wrong, the quoted sentence, the value they stated and your canonical answer. The most-contradicted facts come first.
Competitor sentences stay out
A claim counts only when its sentence, or the nearest earlier sentence naming a tracked brand, is about you. Sentences about a tracked competitor are skipped, so a tracked rival's price is not scored as your mistake.
Accuracy score by engine
One score, supported claims divided by supported plus contradicted, and the same score per engine, so you can see which assistant misstates your facts most often.
Facts suggested from your own pages
Claude Haiku reads up to four of your key pages and proposes up to 20 facts, each with an evidence quote and source URL. Nothing is saved until you accept it.
Three steps, start to finish
Step 1: Add the facts that matter
Type them in or click Suggest from site. Good first facts: a plan price, founding year, headquarters, whether you have a free plan, headcount and your docs URL.
Step 2: Every new answer gets checked
When a scan of your tracked prompts lands, its answers are compared with your library. Re-check answers reruns the last 30 days on demand, at no credit cost.
Step 3: Correct what engines get wrong
Start with contradicted critical facts. Each one shows the engine, the sentence and the value it stated, and later scans add to its right and wrong counts as engines catch up.
The specifics
Engines, data sources, cadence and plan limits, stated as the product works today.
- Where it lives
- AI Search > AI Accuracy in the app, plus the get_ai_accuracy and add_brand_facts tools in Agent M and the MCP server
- Answers checked
- Every answer from your tracked-prompt scans is checked as it lands. Re-check answers and fact edits rebuild the verdicts from the last 30 days. Perplexity scans daily; ChatGPT, Gemini and Claude weekly, or daily on Enterprise
- Fact types
- Number, percentage, date, yes/no, URL and text. Up to 60 facts per project, each with up to 12 anchor words and a critical flag; numbers also take a unit and tolerance
- Verdicts
- Supported, contradicted or needs review. Text facts never count as contradicted. Score is supported divided by supported plus contradicted
- Cost
- Checks and re-checks use no credits and no model call. Fact suggestions make one Claude Haiku call and are rate limited
- Plans
- Included on Starter, Pro, Advanced and Enterprise
Common questions
How does Measure decide whether an AI answer gets my brand facts wrong?
Measure splits each answer into sentences, keeps the ones about your brand that mention a fact by its label or anchor words, and reads the stated value. Numbers and percentages are compared with unit scaling and your tolerance, dates by year and month, yes/no facts by negation and URLs by address. No model is asked to judge, so the same answer always gets the same verdict, and checking uses no credits.
Does the AI accuracy monitor catch every false statement about my brand?
No. It checks the answers Measure collects for your tracked prompts against the facts in your library, and a re-check covers the last 30 days. A statement is only caught when its sentence uses one of the fact's labels or anchor words, so heavy paraphrases can be missed, and topics you have not added a fact for are not checked at all.
What happens when an AI answer describes a text fact differently?
Text facts, such as a headquarters city or a plan name, count as supported when the answer contains your phrase or is a close character-level match. When they differ, Measure marks the claim as needs review instead of wrong, because a phrase can be reworded many ways. Only numbers, percentages, dates, yes/no facts and URLs are marked contradicted automatically.
Will a competitor's pricing in the same answer count against my brand?
Not when the competitor is tracked. A sentence counts toward your accuracy score only when it, or the nearest earlier sentence naming a tracked brand (up to four sentences back), is about your brand, so sentences about a tracked competitor are skipped. A sentence with no brand named nearby is only used when the answer mentions no other tracked brand. A sentence that names you and a rival together is read as being about you, and brands you do not track are not recognized, so add your main rivals as tracked competitors.
Do I have to enter every brand fact by hand?
No. Suggest from site has Claude Haiku read up to four of your key pages, or your homepage, pricing and about pages, and propose up to 20 facts, each with the sentence it came from. Every suggestion starts ticked, so untick anything that is not true before adding the rest. Agent M can also draft facts from your pages and save them with the add_brand_facts tool. It is instructed to confirm values with you first, and every fact stays editable on the AI Accuracy page.
Know where your brand stands in AI answers.
Plans start at $89 a month. Add your brand, choose the prompts your buyers ask, and see who AI engines recommend.