How to Check If AI Is Telling the Truth About Your Brand with Qwairy’s MCP

An engine calling you an entry-level option when you sell to enterprises is not failing to mention you. It is mentioning you and framing you wrong, which no visibility number surfaces. Here you read the framing itself, check the perception score against the wording it summarises, and leave with a correction plan tied to real sources.

The GEO problem it solves

The instinct when a description is wrong is to restate the positioning louder: new homepage copy, a new deck, a campaign. That misses, because an AI description is assembled from third-party sources the engine already trusts.

The second problem is the headline metric. get_brand_perception returns one stored snapshot: sentiment, alignment, consistency and factual-alignment scores, plus a SWOT of bare label strings with no supporting text and no trace back to any answer. Enough to tell you a judgement was made, not enough to audit it. So this workflow pulls the verbatim wording afterwards and lets it win when the two disagree: a healthy alignment score sitting above contradictory descriptors is a reason to trust the descriptors.

How the workflow breaks down

1. Read the current snapshot. get_brand_perception gives the four scores and the four SWOT lists. If the snapshot comes back null and the lists empty, no analysis has been stored: skip to step three and build the picture from answers alone.

2. Check whether a trend exists. get_perception_history returns saved snapshots keyed by month and year, each with its own scores and SWOT, plus a total snapshot count. That count is the test: at zero or one there is no trend, so say so and move on.

3. Pull the answers themselves. get_answers takes a start and end date and returns each answer with its prompt text, provider, date and a brand-mentioned flag. It does not filter on that flag, so filter it yourself. If nothing comes back mentioning the brand, that is an absence problem, not a framing one, and this stops.

4. Read the exact wording. get_answer_details returns the full generated text, competitor mentions with position and sentiment, and the source list with a self marker per source. Build a descriptor inventory quoted verbatim, per provider, with the cited sources alongside.

5. Compare against the positioning you supply, and rank the gaps.

What to do with the output

A descriptor earns work when it clears two bars: it appears in answers from at least two providers, and at least one of those answers returns a source you can identify. Both matter: the object of the work is the source.

Below that line, be strict. A single-provider descriptor that still traces to a source is a watch item, re-checked at the next monthly snapshot. A descriptor with no cited source behind any occurrence gets no budget: there is nothing to go after.

One reading to refuse: a later snapshot moving is not proof the correction worked. Check whether the descriptor itself changed in the retrieved wording.

The tie-break: when two descriptors clear both bars and you can fund one, take the one contradicting a claim you sell. A contradiction gets repeated back at you in a buying conversation; an absence just sits there.

The Complete Prompt

I want to audit whether AI engines describe my brand the way I position it. List the brands I monitor and ask which to use if there are several. Before step five, ask me for my positioning statement and the two or three attributes I sell on: the gap analysis is meaningless without them and no tool holds them.

1. Run get_brand_perception. Show the four scores and the four SWOT lists, and say plainly that the SWOT items are labels with no supporting text. If the snapshot is null and the lists are empty, tell me no analysis has been stored and go straight to step three.
2. Run get_perception_history. Report the total snapshot count first. If it is zero or one, say there is no trend and move on. Otherwise show how each score moved across the monthly snapshots and name the largest shift.
3. Run get_answers over the last 30 days. Filter to answers mentioning my brand yourself, since the tool does not. Select a representative set across providers, favouring prompts where they disagree. If none mentions my brand, stop and tell me this is an absence problem, not a framing one.
4. Run get_answer_details on that set. Quote the exact wording each provider uses, then build a descriptor inventory, per provider, with the cited sources behind each descriptor.
5. Rank the gaps against my positioning. Act only on descriptors appearing across two or more providers with at least one identifiable source. Single-provider descriptors that still trace to a source go on a watch list. Recommend no spend where no source sits behind the descriptor, and where two items qualify, put the one contradicting a claim I sell first. For each funded gap give the source evidence, the hypothesis labelled as one, a correction action, and the wording to change.