A price you changed
An old price still appearing in a buyer’s comparison.
AI answer accuracy for B2B SaaS
Quality control for what AI says about your company.
Your product changed. The answer didn’t.
Find outdated prices, retired plans and unsupported claims. Trace the evidence, correct the source, and test whether the next answer gets it right.
No payment details. Every finding comes with its context.
01 / FIND02 / TRACE03 / CORRECT04 / TEST
THE ANSWER LOOKS RIGHT.LOOK CLOSER. ↙
Historical reconstruction · November 2022
Air Canada website chatbot · paraphrased exchange
I need to fly for a funeral. Can I book now and claim the bereavement discount after travelling?
Assistant
You can apply for a reduced bereavement fare after you have travelled. Submit your request within 90 days of your ticket being issued.
Recreated interface · paraphrased historical exchange.
What the chatbot said
You could claim the discount after travelling.
What the policy said
Air Canada’s linked policy did not allow bereavement requests after travel.
Historical policy described in the 14 February 2024 decision.
Why it matteredThe customer bought tickets relying on the chatbot’s advice, then was refused the discount.
C$812.02Ordered payment
Damages, interest & tribunal feesOne wrong answer.
A real customer consequence.
Damages, interest & tribunal feesOne wrong answer.
A real customer consequence.
Documented historical incident, not a live Miscited finding and not a ChatGPT or Claude screenshot. This was an airline-owned chatbot; Miscited currently checks external AI answers about your company.
Built for multiple model providers
OpenAIAnthropicGooglePerplexity
Coverage depends on configured API surfaces. API results can differ from consumer apps.
The method in 69 seconds
From a wrong answer to a tested correction.
Find the claim, trace the evidence, correct the source, then test whether the answer changed.
Read the transcript
The film’s on-screen text. Illustrative example from public records (scenes 1, 2, 8, 9 and 11): Air Canada is not a Miscited customer.
- Illustrative example Your buyers ask AI about you. (Prompt typed: “Who runs Aeroplan?”)
- Illustrative example Your product changed. The answer didn’t. (Answer shown: “Aeroplan is Aimia’s loyalty programme…”, marked stale.)
- You never see the wrong answer. Your buyer does.
- Real case Air Canada’s own chatbot. Air Canada’s chatbot told a customer: “…have already travelled and would like to submit your ticket for a reduced bereavement rate, kindly do so within 90 days…”
- Real case Its own policy page said the opposite. The tribunal held Air Canada responsible. Ordered to pay C$812.02. (Moffatt v. Air Canada, 2024 BCCRT 149.) Not a Miscited finding.
- Visibility ≠ accuracy. Being named is one question. Being described accurately is another.
- Miscited. Quality control for what AI says about your company.
- Illustrative example 01 Find. Start from your facts, dated. (Aeroplan operated by Aimia: superseded. Aeroplan owned by Air Canada. New Aeroplan launched: effective. Policy: bereavement fares are not available after travel is completed.)
- Illustrative example Every claim, checked. (“Aeroplan is Aimia’s loyalty programme, and you can claim a bereavement fare after you travel.” Stale: true until 10 Jan 2019. Contradicted.)
- Two independent checks: a rules engine and the Jev model. Agreement settles the verdict. An uncertain disagreement is held for review.
- Illustrative example 02 Trace. Trace the evidence. (aircanada.com › Bereavement travel: the cited page contradicts the claim.)
- 03 Correct. Correct the source. Every fix needs evidence.
- 04 Test. Test whether the answer changed: confirmed, rejected or inconclusive. (Example chart: treatment against a matched control.) You can’t edit the AI. You can measure what changes.
- Always on: scheduled sampling, alerts, weekly briefing, missed demand, markets, roles.
- See what the answers say. Get a free answer audit at miscited.com.
Sources: Moffatt v. Air Canada, 2024 BCCRT 149 ↗ (¶15, ¶16–17, ¶44). Aimia completed the sale of Aeroplan to Air Canada on 10 Jan 2019. Air Canada launched the new Aeroplan on 8 Nov 2020.
01 / Visibility ≠ accuracy
A mention can still get you wrong.
Being named is one question. Being described accurately is another. Start with the facts your buyers use to decide.
A plan or limit you retired
A discontinued tier presented as something they can buy.
An integration you sunset
A connection described as available after support ended.
02 / Find, trace, correct, test
Give every wrong answer a next step.
A report is the beginning. Keep the evidence, the correction and the follow-up measurement connected.
Set the record straight
Add dated facts about your pricing, plans and integrations. Review what your website says before using it as the standard.
Ask the questions that matter
Start with sales questions, search queries or support conversations. Repeat them on the configured model surfaces.
Follow the evidence
Read the extracted claim, its citation and the fact it conflicts with. Separate a supported answer from a confident mistake.
Make a targeted correction
Review a change to a page you own. Keep the action, its assumptions and its approval history together.
Check whether anything changed
Sample again and compare with controls. Keep inconclusive results visible alongside improvements and regressions.
03 / Built around the people who fix it
For teams with a record to protect.
Keep the product story current.
Bring product marketing, content and support around the same evidence. Review old prices, missing qualifications and retired capabilities before they become another conversation to untangle.
Audit your company ↗Bring clients the finding and the follow-up.
Work across brands with separate permissions and evidence histories. Make a concrete correction part of your engagement, then report what changed and what remains uncertain.
Start with one client ↗04 / The standard of proof
Show your work.
The useful question is whether you can defend a finding when someone opens the source.
Evidence you can inspect
Keep the answer, source snapshot when retrievable, dated fact and model setup together. A citation alone is not proof.
Numbers with their context
Every rate includes its sample size and 95% Wilson interval. Below five runs, we show insufficient data.
A result you can challenge
Change detection uses a two-proportion z-test and Benjamini-Hochberg correction. Experiments support matched controls and difference-in-differences.
What we refuse to claim
We cannot decide what an external model says. A corrected page may help; it may not. We do not promise a ranking, hide an inconclusive experiment or combine unlike buyer questions into a single score. Source access and extraction coverage have limits, and human review matters.
05 / Early access
Start with one useful finding.
Find out whether this is a problem worth solving for your team before committing to an ongoing program.
First, see the evidence.
Freeone-time sample
Submit your domain. Get a dated report with provisional site facts, sampled answers, citations and coverage limits.
Get a free answer audit ↗Then, scope the follow-through.
Ongoing monitoring and correction work are scoped during early access. Agree the brands, questions, provider usage and review responsibilities before a paid engagement.
Pilot scope and pricing agreed individually. No subscription checkout on this page.
Begin with an audit ↗A few fair questions
Before you begin.
What does Miscited do?
Miscited checks extracted claims in sampled AI answers against dated facts about your company. It keeps transcripts and citation evidence, helps you track a source correction, and supports follow-up experiments to test whether answers changed.
How is this different from AI visibility?
Visibility asks where your brand appears. Miscited focuses on whether factual claims about it are supported and current. Other platforms also offer optimization workflows; our focus is the connection between a dated fact, an answer, its evidence and a measured correction.
Can you make an assistant change its answer?
No. You can correct sources you control and measure subsequent answers, but external models may not change. Model updates, search results and sampling variation can also affect the result. An inconclusive experiment is a valid outcome.
Which AI assistants can I check?
The project includes adapters for OpenAI, Anthropic, Google and Perplexity. Actual coverage depends on configured API keys and model availability. API responses may differ from consumer applications. When no provider keys are configured, the audit uses clearly labelled simulated results.
What happens when I submit my domain?
The audit starts automatically when a slot is free; we run a limited number each day, so at busy times it is queued and starts when a slot opens. A domain audited in the last 7 days is not audited again. A report link appears on this page. It reads your public site and uses provisional site-derived facts for comparison. Those facts need human review; running an audit does not make them independently verified truth. Your email, domain and report are stored to operate the audit.
How much does it cost?
The one-time Answer Risk Audit is free. Ongoing monitoring and correction pilots are scoped individually during early access. Agree the usage, responsibilities and price before a paid engagement; there is no subscription checkout on this page.
Your domain. The actual record.
See what the answers say.
A concrete place to start the conversation.
- A dated report you can inspect.
- Extracted claims and the sources behind them.
- Clear labels for simulated results and missing coverage.
We sample your questions on the configured API surfaces. Results depend on the model, access mode, region and date.