Our AI-Written Competitor Comparisons Were Lying About Our Own Product

The Failure Mode Is Not What You Expect

Everyone worries that an AI writing about your product will overclaim. It will invent an integration, promise a feature you do not have, or hallucinate a pricing tier. That is the risk people plan for, and it is a real one.

The failure we actually hit was the opposite, and it was much harder to catch.

We generated a cluster of competitor "alternatives" articles. Miro alternatives, Excalidraw alternatives, FigJam alternatives, Milanote, tldraw, Apple Freeform, Whimsical, Mural, Lucidspark. Standard comparison content. The drafts read well. The structure was right, the competitor descriptions were accurate, the tone was reasonable and not overly promotional.

Then we read them against the actual product, and found that across eleven posts, the articles were confidently and repeatedly understating what we ship.

What the Articles Claimed

A representative sample of what we had published about ourselves:

• The Miro alternatives post framed OmniCanvas as "not a workshop host" that "lacks real-time multiplayer"
• The FigJam alternatives post described it as a solo, private notebook
• The Lucidspark post drew a line between "personal tools" and "team suites" and put us firmly in the personal column
• The Mural post advised readers to pick us only "if you do not need live facilitation"
• The Excalidraw and tldraw posts routed every collaboration requirement to Miro and FigJam
• The Milanote post did not list us among collaborative creative canvases at all

OmniCanvas has real-time collaboration. Live cursors, permissions, invites, up to 25 simultaneous collaborators.

We had written, published, and were actively driving search traffic to eleven articles telling prospective users that we could not do the thing we do.

Why This Happens

The mechanism is straightforward once you see it, and it generalizes to any product.

A language model writing about Miro is on solid ground. Miro is enormous, well documented, discussed in thousands of articles, and its capabilities are effectively common knowledge. The model has a strong, accurate prior.

A language model writing about your product has almost nothing. You are small. Your docs may not be in the training data. What the model has instead is a category prior: a small independent canvas app made by a tiny team. And it knows what those are usually like. They are usually single-player. They usually lack real-time sync, because real-time sync is hard and expensive and small teams usually skip it.

So it fills the gap with the category average. And the category average, for a small app, is "no multiplayer."

This produces text that is fluent, plausible, appropriately hedged, and wrong. It does not read like a hallucination. It reads like sober analysis. That is exactly what makes it dangerous — an overclaim trips your internal alarm immediately, and a modest, reasonable-sounding underclaim slides right past.

The Second Version of the Same Problem

We found the mirror image in a different cluster, and it is worth describing because the correct fix went the other direction.

Our AI meeting-notes posts repeatedly described OmniCanvas encryption as blanket and always-on. Phrases like "AES-256 zero-knowledge encryption," "end-to-end encrypted," "encrypted storage," "encrypted notes," used as general descriptions of the product.

That is not how it works. Encryption is opt-in, per note or per folder. A user who reads "your notes are end-to-end encrypted," assumes it applies to everything, and stores something sensitive in a note they never explicitly encrypted has been actively misled by us.

Same root cause, opposite direction. The model knew that privacy-focused note apps in this category tend to advertise blanket encryption, so it wrote the category convention rather than our specific behavior.

Underclaiming collaboration costs you signups. Overclaiming encryption is a trust and potentially a legal problem. The second one is much worse.

What We Actually Fixed

We corrected eleven posts across three files. The pattern of each fix:

Note what the corrections are not. We did not turn "OmniCanvas cannot do collaboration" into "OmniCanvas is better than Miro at collaboration." Miro genuinely is better at facilitated enterprise workshops, and the corrected posts say so. The goal was accuracy in both directions, and the honest version of the Miro comparison still recommends Miro for a large chunk of readers.

What We Changed About the Process

Every claim about our own product needs a source in the codebase or the docs. Competitor claims can come from public material. Self-claims cannot come from the model's priors, because the model does not have any.

Give the model a capability sheet, not a product name. If the draft is generated from an explicit, current list of what ships — including scope and limits like "encryption is opt-in per note" — the category prior has nothing to fill in.

Read for what is missing, not just what is wrong. Every one of these errors was an omission or a hedge. Proofreading catches false statements. It does not naturally catch a true-sounding sentence that quietly excludes you from a list you belong on.

Audit the highest-traffic posts first. One of the affected articles was drawing well over a thousand sessions a month and producing zero signups. We had assumed the CTA was weak. Part of the problem was that the body text told readers we could not do what they came to do.

The Thing Worth Remembering

If you use AI to write about your own product, it will describe the category you appear to belong to, not the product you actually built. Where you are typical, that works. Where you are unusual — where you did the hard thing a small team normally skips — it will quietly erase your best differentiator and sound perfectly reasonable doing it.

Check the places where you are exceptional. Those are exactly the sentences that will be wrong.

If you want to see the collaboration we spent eleven articles denying, open a shared canvas and send someone the link.

30 days free. No credit card required.