DiscoveryHQ
Loading…
DiscoveryHQ
Loading…
Responsible AI
This whole product is built on the output of AI models, so how we handle that output is not a policy sitting next to the product — it is the product. Here is what we commit to, what we refuse to do, and the limits we would rather tell you about now than have you discover later.
Our method
There is one rule underneath all of this: if we did not observe it, we do not print it. Everything below follows from that.
Nothing here is modelled, predicted or estimated. We ask ChatGPT and Gemini the questions your customers ask, with web search turned on, through the AI companies' own official services, and we write down what came back.
Ask an assistant the same question twice and you can get two different answers. So every tracked question is asked 3 times on each engine, and the count you see is the count across all of them.
You will read “recommended in 2 of 9 checks”, because that is a fact you can audit. There is no confidence percentage anywhere in the product, because no such number exists to report.
Every number opens onto the answers behind it: the question we asked, the engine that answered, and the full text it replied with. If you cannot check a figure yourself, we should not be printing it.
A score you cannot take apart is just an opinion with a number on it. Every figure in DiscoveryHQ has these four layers underneath it, and you can walk down all of them.
One number, 0 to 100. Tap it and it stops being a number.
The real customer questions we asked on your behalf, listed one by one.
ChatGPT and Gemini, 3 tries each, kept separate rather than blended into an average.
The full reply, including who it named instead of you and the reason it gave.
The free public check-up is 3 questions across 2 engines, 3 tries each — 18 separate checks, in about one to two minutes. Every one of them is stored and openable.
What we refuse
A product about AI answers could invent one and almost nobody would catch it. These are the four places we could have cheated and did not.
Every result in your report came back from a live call to a real assistant. We never generate a plausible-looking answer to fill a gap in a report.
We never write text ourselves and present it as something a model said. A stored answer is the model's answer, exactly as it arrived.
No made-up testimonials, no customer numbers, no logos of companies that never used this, no awards nobody gave us. If it is not real, it is not on the site.
Two scores exist: your AI Visibility Score and your Website Score. We will not invent a friendlier third one because a page looks empty without it.
Handling model output
Trusting model output is how AI products get things badly wrong. We assume any answer might contain something unexpected, and we check before we use it.
A model can reply with anything at all. Before an answer is counted, stored or shown to you, we pull out only the specific pieces we asked for — the names it mentioned, where in the answer they appeared, the reason it gave — and check each piece against a strict shape. Anything that does not fit is dropped rather than guessed at.
Free text from a model never travels any further into the product than the moment it arrives. Only the checked pieces move on.
When we read a website page by page, that page is somebody else’s code and we treat it the same way: read it, take out the parts we need, and never let the raw page decide what happens next. The in-app check-up covers up to 100 pages.
Your results are stored in Postgres with row-level security, so one business’s data cannot be read from another business’s account. The Security page covers that in full.
Who decides
AI writes the suggestion. A human reads it, judges it, and chooses whether it goes anywhere near their business.
When a check-up finds something missing, we write the wording for you — the answer to a question your site never answers, the details AI could not find, the small code label that tells AI what your business is. Then it is yours to read, change, or throw away.
Nothing is ever applied to your website automatically. We hold no keys to your site and make no edits to it. You paste in what you agree with, and a change only counts as done when another check-up confirms it — we never mark something fixed because we suggested it.
You know things about your business that no model does — what you actually offer, what you will not promise, what your lawyer would say. A fix that skipped you would be a fix nobody checked.
It also means a mistake in generated wording stays a suggestion. It cannot become a live page on your site without a person deciding it should.
Where we draw the line
There is a real difference between being easier to understand and being easier to believe. We only do the first one.
They take true things about your business that AI could not find or could not read, and make them plain: what you sell, where you work, what it costs, who it is for, and the answers to the questions people actually ask. A real business, described clearly.
Invent credentials, claim services you do not offer, fake reviews or awards, or dress up a business as something it is not. Making a false business easier to recommend is the one thing this product exists to not do.
This is not only a principle, it is a rule you agree to. Using DiscoveryHQ to help an assistant recommend something untrue about a business is a breach of our Acceptable Use Policy, and we will close an account over it.
Limits
Every measurement tool has edges. Ours are easier to work with once you know where they are, so here they are.
The same question can get a different answer tomorrow. That is how these systems work, and it is exactly why we ask each question several times instead of once.
It tells you what happened when we asked, not a permanent verdict on your business. Treat a single run as one reading, and a run after a fix as the thing that confirms it.
An assistant can describe your business inaccurately, miss you entirely, or lean towards the names it has seen most. We report what it said. We do not defend it.
ChatGPT and Gemini, and only those two. Perplexity and Google's AI Overviews are not covered, so nothing here should be read as speaking for them.
It is not a guarantee of ranking, traffic or revenue. It measures how often AI recommends you for the questions you chose to track — which is useful, and which is all it is.
None of this makes the numbers less useful. It means you read them as what they are: a record of what two assistants said about you, over a set of questions you chose, at the moment you asked. That is a far better starting point than a guess — and it is not a crystal ball.
Data and models
A check-up means sending questions to OpenAI and Google. This is exactly what is in those questions.
The AI providers are OpenAI and Google, reached through their own official services with web search turned on. Beyond them, the list of companies that touch your data is short and every one of them is described by the job it does in the Privacy Policy.
If something is wrong
We report the answer, we do not endorse it. When an assistant gets your business wrong, that is worth looking at properly.
Write to support@dicoveryhq.com and tell us which check-up it was and what the assistant got wrong. Because we keep every answer word for word, we can open the exact reply and see what came back rather than take a guess at it.
What we can do
Show you the stored answer, check whether we counted it correctly, fix it if we got it wrong, and often tell you what on the open web the assistant was reading.
What we cannot do
Reach inside ChatGPT or Gemini and change their answers. Nobody outside those companies can. What usually shifts an answer is fixing what the web says about you, which is the work this product is for.
DiscoveryHQ is developed and operated by Loom Labs AI LLC. One inbox, and a person reads it.
A free check-up takes a minute or two. No account needed to run it, and a free account saves the report.