AI is already answering questions about your company. Who checks the answers?
An analyst wants to know who leads your company. A journalist checks where you are based. A pharmacist asks what your product is for. Instead of opening your website, they ask an AI and get an answer in seconds that sounds certain, whether or not it matches your annual report.
The Seravel AI Answer Audit measures what engines such as ChatGPT, Claude, Perplexity and Google Gemini say about your company and your products, against your own verified sources. Rates instead of screenshots, a traced source for every error, an owner for every fix. Nobody can make a model answer correctly; the audit shows what you can steer yourself and whether it works. Independent: I sell no implementation.
Why a screenshot proves nothing
AI models are inherently probabilistic and stateful. Relying on a single snapshot leaves your digital compliance and brand reputation unmeasured.
Ask twice, get a different answer
When drug questions were entered into ChatGPT a second time, only 3 of 12 answers came back identical[cite: 1]. In ChatGPT and Google’s AI, repeated prompts for product recommendations returned the same list of brands less than once in a hundred times[cite: 2]. A single answer is an anecdote.
Cited isn’t the same as used
In a test with 50 drug questions, most of the wrong or partly correct ChatGPT answers gave no references[cite: 1]. Where AI search does cite, a 2026 preprint across ChatGPT, Google and Perplexity shows that being cited and shaping the answer are two different things. Whether your own pages shape the answer has to be measured.
Today’s answer is not next quarter’s
The same AI service can change within months: in one study, GPT-4 identified prime numbers correctly 84% of the time in March 2023 and 51% in June. A measurement needs a date, a model and a baseline.
You don’t control the model. You control these three things.
While provider algorithms remain opaque, your underlying digital evidence, compliance documentation, and structured web presence are entirely within your operational reach.
Your content and its structure
The audit shows which questions AI gets wrong and which sources it relies on. Part of that you can influence: your own pages, FAQs, structured data and product information, improved exactly where the answers go wrong.
Evidence for Regulatory and Compliance
A record of who checked what and when, which critical deviation was found and which measure followed. Documented diligence instead of screenshots in an inbox.
Visibility against competitors
Whether you are named in general questions, whether your key messages arrive, and who is recommended instead. You steer through your content; the next measurement shows whether the answers have changed.
Three questions the audit answers
A rigorous diagnostic framework designed to establish absolute clarity on model outputs and pinpoint root causes for remediation.
What is being said?
I ask the engines the questions your stakeholders ask, repeatedly and reproducibly.
Is it true?
Every answer is checked against the reference you provide. The result is a hit rate, not an impression.
Where does the error come from?
I trace wrong statements to their likely source – directory, database, press release, your own website – and name who owns the fix on your side.
One question, five runs, one outdated source
Concrete transparency into how an outdated directory listing can override your official product pages inside an AI response.
Illustration with a fictional company
- Question
- Where is Muster Medical AG headquartered, and who leads it?
- Engine, language, date
- Engine A (API), English, October 2026
- Five runs
- 1 of 5 correct
- Cited source
- A business directory entry, last updated before the company moved.
- Not cited
- The company’s own page with the current address and management.
- Diagnosis
- The outdated directory entry outweighs the company’s own page. Owner of the fix: Communications, via the directory’s correction process. Then measure again.
Every finding states engine, model, language, date and number of runs. I report rates, never single answers, and archive each run as the baseline for the next measurement.
Five steps, two weeks
A transparent, repeatable workflow moving from truth definition to automated execution, diagnostic analysis, and board-level briefing.
-
Days 1–4
Define the truth
Together we define 12 to 16 questions per product that customers, investors, journalists or healthcare professionals would ask an AI. For each one you provide the verified reference answer – from the annual report, your website or the product information. On request, the tool proposes statements with their source passage and you confirm them. Only facts confirmed by a person count.
-
Days 5–7
Measure
Each question runs across the agreed engines in your languages, five runs each, with and without web search where the engine supports it. An AI model from a different provider than the one that answered breaks each answer into individual statements and checks them against your reference: confirmed, wrong, unsupported or key message missing, rated by error type and severity. Per question: visibility rate, accuracy rate, spread, cited sources, archived baseline.
-
Days 8–10
Review and diagnose
I review a sample of the ratings; safety-relevant findings stay provisional until a person has confirmed them. I trace wrong and missing answers to their likely source: outdated directory entries, conflicting product data, missing structured data, press texts, Wikipedia. Each finding is rated by its possible impact on reputation, revenue or compliance. Whether a deviation is compliance-relevant is for your own teams to decide.
-
Days 11–12
Translate
The report and a 90-day plan: measures, an owner per source, an approval process and a measurement rhythm.
-
Days 13–14
Brief and hand over
A 60-minute briefing for the executive board or the board of directors, and the decision: monitoring, a fractional mandate or neither.
Beta: the measurement tool is being extended engine by engine. Some engines run automatically, others through a manual audit that is assessed the same way. Which engines and languages your audit covers is confirmed in writing before we start. Answers via the providers’ interfaces can differ from what the consumer apps show; where agreed, I add manual spot checks in the apps, and every result says which route it came from.
Seravel: the tool behind the audit
Built from the ground up for rigorous, repeatable verification across complex digital ecosystems.
What it measures
Four areas: company facts, product information, key messages and visibility on recommendation questions. Every question five times, with and without web search, through the official APIs; manual spot checks in the chat apps where it matters.
Who rates
A model from a different provider than the one that answered breaks each answer into statements: confirmed, wrong, unsupported, key message missing. A person reviews the sample; safety-relevant findings stay provisional until confirmed.
Where the data lives
Reference, questions, answers and ratings stored separately per client. Only public, approved information is processed. Only you see the results; findings, reference and method stay with you after the engagement.
Seravel is in beta. For the pilot phase I am looking for organisations in pharma, OTC and MedTech; pilot clients get a preferential price in return for a case study.
What you get
- An audit report of 20 to 30 pages, opening with a two-page summary: the key message in plain language, the three most important findings (safety, accuracy, visibility) with evidence, at most three recommendations with owners, and figures per AI system.
- An evidence file: every archived answer with engine, model, language, date and run, as your baseline.
- A 60-minute briefing for management or the board, with Communications and Legal if you wish.
- A 90-day plan with measures, an owner per source and a measurement rhythm.
For whom
- Pharma, OTC and medtech: where a wrong product statement can become a compliance case. For consumer health and OTC there is a dedicated page.
- Brands under competitive pressure: when AI recommends a competitor instead of you.
Start free. Fixed prices after.
Transparent, modular pricing designed to scale from an initial risk snapshot to comprehensive multi-product monitoring.
Snapshot
Free
5 questions, 1 engine, 1 page. Within 48 hours. For medicines: the safety statements compared with your package leaflet.
In return: 15 minutes to walk through the result.
Audit
Fixed price
One product: 12–16 questions, 2 engines, 2 languages. Report and briefing.
One company or one product area.
Audit Regulated
Fixed price
Several products, 12–16 questions each, 4 engines, 3 languages. Findings prioritised by compliance risk, board briefing.
Pharma, OTC, medtech, group brands.
Monitoring
Every 14–28 days
Rerun every 14 to 28 days, deviation report, alert on new false statements. Each rerun shows whether the answers have changed since your measures; model updates play a part too, so it is not proof of cause.
After the audit, for up to six months – then your team takes over.
Scope and fixed price are confirmed on the scoping call. All prices in CHF, excluding VAT. The audit fee is credited against a fractional mandate (minimum 3 months) started within 60 days.
What the audit is not
Absolute clarity on the limits of advisory scope and operational responsibilities.
No implementation
No agency hours. I audit and translate the findings into responsibilities; execution stays with your teams or partners.
No guarantee
Nobody can make a model answer correctly, and AI answers change without notice. The audit shows where the gaps are, which of them you can influence, and whether the answers change.
No medical, regulatory or legal advice
The audit shows deviations from your approved statements. It is no substitute for your Medical, Regulatory or Legal approval.
Frequently asked questions
Can you make the AI answer correctly?
No, and nobody can. AI providers decide how their models answer. What you can influence is your own content, your evidence and your visibility. The audit shows where to act, and the next measurement shows whether the answers have changed.
Why not simply ask the AI ourselves?
Because a single answer proves nothing. The same question can be answered differently on the next attempt. I measure rates across several runs, compare them with your verified reference and archive them as a baseline.
What makes this approach distinctive?
Four things, taken together. The yardstick is your own verified reference, not an outside opinion. Results are rates across repeated runs, each documented with engine, model, language and date and archived as a baseline. Every wrong statement is traced to its likely source and assigned to an owner on your side. And the audit is independent of execution: I sell no implementation, no content and no agency hours, so the recommendations are not tied to follow-up work of that kind. Behind it are thirteen years on the client side of digital platforms, including in a regulated industry.
Do you also sell the implementation?
No. I audit, trace each error to its source and name who owns the fix. Implementation stays with your teams or partners.
What happens with our data?
Questions and reference answers come from your public, approved sources; no personal data of customers or patients. To measure, the questions are sent to the AI providers being tested via their interfaces, and the answers are checked by an AI model from a different provider. Your facts, questions and results are stored in a database in Frankfurt (EU), separated per client at database level. Under their terms, the AI services we query do not use these API requests for training (Anthropic; Google in the paid tier we use). Results are shared only with you, not published and not compared with other companies. Contract under Swiss law.
Are API answers the same as what people see in the apps?
Not necessarily. Apps add their own instructions, memory and location. I measure through the providers' interfaces because it is reproducible, add manual spot checks in the apps where agreed, and every result says which route it came from.
What comes after the audit?
Either monitoring for up to six months, until your team runs the measurement itself – with my tool or with tools of your choice – or a fractional mandate. Both end with a handover.
Sources
- Morath, B. et al. (2024). Performance and risks of ChatGPT used in drug information. European Journal of Hospital Pharmacy, 31(6), first published online 2023. (50 consecutively collected drug-information questions; answers rated by six senior hospital pharmacists; ChatGPT version of 2023.)
- Fishkin, R. & O'Donnell, P. (2026). AIs are highly inconsistent when recommending brands or products. SparkToro; summary in Search Engine Journal. (2,961 runs of 12 brand-recommendation prompts by hundreds of volunteers in ChatGPT, Claude and Google AI Overviews; not peer-reviewed; co-author Patrick O'Donnell works for the AI tracking start-up Gumshoe.ai.)
- Zhang, K., He, X. & Yao, J. (2026). From Citation Selection to Citation Absorption: A Measurement Framework for Generative Engine Optimization Across AI Search Platforms. arXiv preprint 2604.25707. (Descriptive study, 602 prompts; no causal claims.)
- Chen, L., Zaharia, M. & Zou, J. (2024). How Is ChatGPT's Behavior Changing over Time? Harvard Data Science Review. (GPT-3.5 and GPT-4, March vs. June 2023; task: identifying prime numbers, 84.0% vs. 51.1%; not health.)
Request your free snapshot
Five questions about one product or topic, one engine, one page – delivered within 48 hours. For a medicine, the safety statements in the answers are compared with its package leaflet. In return I ask for 15 minutes to walk you through the result.
One short form: company, website, product or topic, languages, contact e-mail.
Or write directly: kevin.lancashire@lancashire-digital.ch