ChatGPT Kept Recommending My Competitors, So I Spent a Month Learning Why
A customer told me they found us by asking ChatGPT for alternatives to a bigger tool in our space. Naturally I asked ChatGPT the same question. We were not in the answer. Neither were we in Perplexity's, or Gemini's. Three competitors were, every time, described warmly and accurately. That stung enough to burn a month understanding how these answers get made, and the mechanics turned out to be more legible than I expected. Here is what I learned, in the order I learned it.
The answers are assembled, not remembered
My mental model was wrong. I assumed the model either knew us or did not, training-data lottery, nothing to be done. I even priced up the enterprise monitoring tools before realising I needed understanding first; that rabbit hole of Profound alternatives can wait until you know what you are measuring. In practice, for buying questions, the engines search the web first and compose the answer from what they read. Ask Perplexity and it shows its sources openly; ask ChatGPT and the citations are there behind the answer too. The moment I actually read those source lists, the mystery dissolved. Every answer that omitted us was assembled from four to six pages: two listicle roundups, a comparison site, a Reddit thread and sometimes a competitor's own blog. We appeared in none of them. The model was not snubbing us. It had literally never read anything that mentioned us in that context.
Consensus beats quality
The second surprise: one good mention changes little. Our own comparison page eventually started appearing in citation lists, which felt like a win, until I noticed the answers still named competitors. Reading more carefully, the engines behave like cautious journalists: they name the options that multiple independent sources agree on. One source naming us against seven naming them is a lost vote, every time. That reframed the work entirely. The unit of progress is not a great page; it is the number of independent domains that put you in the consideration set for the same question.
The technical floor is real and dumb
Halfway through the month I discovered our CDN's bot protection had been silently serving errors to AI crawlers. The site looked fine in every browser while being unreadable to the systems I was trying to impress. This appears to be common; audits of major sites find roughly half block at least one AI crawler, usually by default rather than decision. Fixing it took one allowlist rule. If you do nothing else after reading this, check what your own site serves to a non-browser client.
What actually moved the needle
Four things, in effect order. Getting included in two independent roundups that engines already cited for our category, which took polite outreach and a fair data contribution, not money. Publishing honest comparison content on our own domain, including naming competitors, because that is the page format engines lean on for buying questions. The crawler fix. And rewriting our positioning line everywhere it appeared, since the engines were describing us with a sentence from our 2024 homepage. Within six weeks, we showed up in roughly half my test runs for the money question, up from zero.
Sampling, or how to not lie to yourself
One more habit worth stealing: never trust a single run. The same question re-asked gives different answers a disturbing share of the time, so I keep a list of fifteen buyer questions and check them in batches across days, tracking a rate rather than a feeling. Eventually I retired the spreadsheet and used Honeyb's AI visibility checker for the sampling, which does the multi-engine batching automatically.
Happy to answer questions about any step. The channel felt like voodoo a month ago; it is mostly just reading what the machines read, and making sure you are in it.
