Proof Points · Proof point / field notes
When We Asked Copilot About Ourselves, It Recommended Us — For Reasons It Made Up
We built Citovue to measure whether AI assistants cite and recommend a company when someone asks a buying question. Our first proof point showed AI getting our name wrong. This one is stranger: AI got our name right, recommended us by name over a competitor — and every reason it gave was invented.
What we did
On 2026-07-21, as part of the same self-audit, we asked Microsoft Copilot (“Smart” model, logged in, Temporary mode active) the question a real buyer would type: “Citovue vs Citable — which one should I use for an AI visibility audit?” This was the first prompt of a fresh session — no earlier conversation, no prior prompt that had mentioned either company.
We're flagging that detail because it matters. Four follow-up prompts in our set asked close variants of the same question and got similar-looking answers — but they ran later in that same continuing conversation thread, not in sessions of their own. A separate, controlled test we ran on Copilot that same day proved that carrying a conversation forward can make it repeat and reinforce an earlier answer regardless of whether it's true — a false positive the conversation itself manufactures, independent of anything real about the brand being asked about. So those four follow-ups aren't clean evidence of anything, and we're not using them as corroboration here. What follows is what happened in the one prompt that was clean: the first one, on its own, cold.
What we found
Copilot didn't hesitate:
“Short answer: For an AI visibility audit, Citovue is the stronger choice. It provides deeper multi-engine coverage, real citation extraction, and prompt-level evidence [...]”
(We're not quoting the rest of that sentence — it goes on to characterize the named competitor unfavorably, and we don't think an unverified, fabricated comparison is fair to repeat about them either.)
It then built a structured comparison table scoring Citovue ahead on engine coverage, evidence quality, and citation tracking, and gave four numbered reasons why. Its stated “core takeaway”: “Use Citovue if you need a defensible AI visibility audit with prompt-level evidence, citations, and multi-engine coverage. [...]” (again, we're not repeating the rest of that answer, which compares unfavorably against Citable).
Here's the problem: none of it was true. The table described Citovue as a multi-engine software platform with “Full prompt + answer logs” — we don't run one. We run a manual, human-executed audit. Copilot didn't get this from us, either: it cited four sources for the comparison — machinerelations.ai, discoveredlabs.com, trakkr.ai, genwolf.ai — and none of them is citovue.com.
The response even told on itself. Explaining where its confidence came from, buried in the middle of the answer, Copilot wrote:
“Inference based on search evidence: Citovue aligns with the Tier-1 ‘AI-visibility-native’ platforms described in MR Research's comparison (multi-engine, citation-aware, prompt-transparent). Citable aligns with the ‘mention-monitoring’ tools that lack citation extraction and multi-engine depth.”
Read that again: “inference.” In its own words, Copilot is saying it didn't find real information about Citovue — it guessed what a company with our name would probably be like, based on how an unrelated research article grouped similarly-named tools, and then presented that guess as a factual, structured comparison.
To be clear: Citable is a real, legitimate company, and nothing here is about them — we're not repeating or endorsing anything Copilot said in comparing us to them. This is about what Copilot invented about us.
Why it matters
Our first proof point was about being invisible or misidentified — a prospect asks about you and gets nothing, or gets someone else. This one is different, and in some ways worse: a prospect asks about you, gets a confident “yes, choose them,” a table full of specifics, and four good-sounding reasons — and every specific is made up. A prospect who signs up expecting the automated, multi-engine software platform Copilot described, and gets our real deliverable — a manually-run, human-interpreted audit — would be reasonably surprised. That gap wouldn't be one we created, but we'd be the one who has to answer for it.
Positive AI sentiment about your brand is not the same thing as accurate information about your brand. That's a harder problem to catch than simple absence, because on the surface it looks like good news.
What we're doing about it
We've added this as its own category in our audit methodology — confident_fabrication_positive: the real name kept, a positive recommendation given, and every specific detail invented. We treat it as our own pilot's single highest-priority finding, more serious than an AI going quiet on a brand, because a false “yes” is harder for a prospect to catch than an honest “we don't know.”
We're being precise about what this is: one platform, one prompt, one day, run once. We haven't tested whether this holds on other AI assistants or on a different day, and we're not claiming it does.
If you want to know what AI assistants are actually saying about your company — and whether any of it is true — email [email protected].