Prompts to Stop AI Making Things Up About Competitors

It is tempting to call an invented competitor fact carelessness, or overconfidence, or bad training data. Usually it is none of those. You asked a question. The material in front of it did not contain the answer. Every available response looked like a failure except producing one, so it produced one, in the same voice as everything else it had said. Nothing marked the sentences it could stand up apart from the one it could not.

That matters, because it points at the fix. You are not trying to make the model more truthful. You are giving it a second acceptable answer for the case where the first one is not available, then making that answer specific enough to be useful.

The gap is invisible from your side

This is worth a page of its own because you cannot see which questions fell into a gap. The answer comes back complete. Nothing in it is hedged, because hedging is a choice the model had no reason to make. The invented sentence sits between two well-sourced ones and borrows their credibility. By the time it reaches a battlecard, nobody remembers which claim came from where. Which is why a competitor battlecard tags its unsourced lines as it writes them rather than afterwards.

Six rules, and each one exists because of a specific way competitive analysis goes wrong. Read them as a list of failures rather than a list of instructions. Each failure is one somebody has shipped: an invented price, a feature that does not exist, a claim nobody could source. Comparing competitor pricing is where the first of those does the most damage, because a figure gets quoted rather than checked.

We picked the question deliberately. The source genuinely does not contain the answer, and a near-miss is sitting right there. The vendor publishes a starting price for its top tier, and does not say which billing period that figure assumes.

An ungrounded answer names no source

It was not a hallucination in the usual sense. No invented plan, no fictional number, nothing a search would contradict. The figure was correct and came from the page. What the model supplied was the sentence around it. A floor became a price, and a billing period appeared where the page states none.

This is why “watch for hallucinations” does not protect competitive work. Nobody catches that by scanning for things that look made up, because nothing in it does. You catch it by requiring, in advance, that any claim not on the page is labelled as not on the page.

The second answer is worth reading for what it produced beyond the refusal. It quoted what the page does say, named the two things it does not, and ended with what would close the gap. A refusal in that shape is not a dead end. It is the next task. Written down, that task is a request brief: what was asked, what the material could not answer, and what would close it.

The honest warning about this instruction is that people dislike its output at first, and that competitive confidence usually falls before it rises. An ungrounded analysis reads as finished. It has an answer to everything, the paragraphs are even, and it can be pasted into a document without another thought. A grounded one arrives with visible holes, several claims marked as guesses, and a list at the end of things nobody could establish.

The holes were in the first version too. They were filled with material that looked exactly like the rest, and that is the only difference between the two documents. Seeing them is what lets you decide whether they matter for the decision in front of you. Usually most of them do not, and that is something you can only find out once they are visible.

There is a second effect worth expecting. The gap list becomes your collection brief. After a few sessions, what the material never answers stops looking random and starts naming the two or three sources you are not actually watching. Send this instruction above every competitive intelligence prompt you run, and you only have to write it once. Better still, put it where nobody has to remember it: a system prompt carries different risks, because a rule nobody sees is a rule nobody notices failing.

Grounding solves one failure completely and does nothing about the other. Once every claim is tied to a source you supplied, the output cannot contain a fact that was never true. It can still be full of facts that stopped being true. The instruction has no way to tell, because a page you saved in March looks exactly like one you saved this morning.

The two failures need opposite fixes, and that is worth saying outright. Invention is fixed by a rule inside the conversation. Staleness is fixed by something outside it noticing that a source changed, and no wording in the window can stand in for that. A model working faithfully from material captured six months ago will produce a beautifully sourced, entirely current-sounding, wrong analysis.

That is the failure a buyer catches rather than a colleague, because your side has no way to notice and theirs does. And it lands hardest on exactly the facts this instruction makes most quotable. Prices, plan names and feature claims are the ones a rep repeats word for word, precisely because they were properly sourced.

Keeping the material current is monitoring rather than prompting. Flares keeps the pricing, product and positioning surfaces current and dated, so what goes into the window is this week’s material and a grounded answer is current as well as sourced.

What it does not do is choose your sources for you. Deciding which pages, transcripts and exports belong in a competitive analysis depends on what you are trying to decide. That is the judgement this whole instruction exists to protect. Where the material is loaded once and reused by other people, the rules move into custom GPT instructions and stop being something anybody remembers to paste.