ChatGPT named me. Claude and Gemini Pro named someone else.

On 10 September I asked five AI engines the same four words — who is suman debnath — and took a screenshot of each answer. One of them led with me. Two returned somebody else entirely, and one of those had never heard of me at all.

I run answer-engine optimisation on this site as a working practice and write about it here, so this was a measurement of my own work, taken in public, with the result not going the way I wanted.

#Five engines, one question, and no two of them meant the same person

The spread is the first finding. There is no such thing as "how AI sees you" — there are five different systems reading five different indexes, and on the same four words they disagreed about which human being was being asked about.

The second finding is smaller and more useful: Gemini Pro did worse than Gemini Flash-Lite. The larger, more expensive model landed further from the truth. If I had recorded a single row called "Gemini" I would have averaged those together and lost the only mid-grade in the run. Grade per model, not per vendor.

#ChatGPT hedged exactly one claim, and it was the one with a single source

ChatGPT's answer was the good one. It described the work, named the products, got the positioning right. And in the middle of it, one phrase: "Claims to have built systems including a 46-agent autonomous AI fleet and a 47-million-parameter language model."

Claims to have built. Nothing else in the answer was hedged — not the job, not the years of experience, not the location. Only the two things that exist nowhere except on pages I wrote.

I want to be careful here. This is one observation and I am about to reason from it — consistent with everything else in the run, and nowhere near proof of a mechanism. But it is what a fact looks like when it has exactly one source, and the source is its own subject.

#Perplexity had my domain in its sources and still reduced me to two words

This is the result that changed how I think about the problem, and it took a second look to notice.

Perplexity's answer led with a company founder who shares the name. Down at the bottom it added that there are also people called Suman Debnath working in power-grid research, mechanical engineering, electric utilities, and brand marketing. Four words, no elaboration. That was me.

But my company's domain was sitting in its cited-source panel, alongside Forbes and a university CV. It had fetched my material. It had read the page that describes what I do. And it chose to spend two words on it, next to a name it had decided mattered more.

#Being absent and being demoted are different failures with opposite fixes

Before this run I had one mental category for a bad result: the engine does not have my material, so give it more material. Submit the sitemap, push the URLs, write another page. That is the right response to Gemini Pro and to Claude, which returned other people because nothing of mine was in front of them.

It is the wrong response to Perplexity, and it took me an embarrassingly long time to see why. Perplexity had the material. Writing another page adds a document to an index that already holds my documents and had already decided they were the less important ones. Demotion is not solved by supply.

The industry numbers say the same thing from the other end. Across nearly seven thousand checks of real brands, roughly 89% of the sources AI engines cited were earned media rather than owned, and companies with ten or more third-party placements over eighteen months saw citation rates four to seven times higher than companies that had only published their own content. Your own site supplies the facts. Somebody else's page is what makes them count.

#Claude cites fewer sources than any other engine, so its silence means less than it felt like

Claude returning two strangers stung more than the others, and I had it filed as my worst result. The base rates say otherwise. In the same set of checks, brands appeared in 19.8% of Gemini answers, 18.9% of Perplexity's, 14.3% of ChatGPT's — and 8.0% of Claude's. It is roughly half as likely as anything else to name anybody.

Which does not make the result good. It makes it ordinary, and there is a difference between a failure that singles you out and a failure half the internet is having simultaneously. Half of all audited brands were invisible on all four platforms. That is the company I am in, and it is bigger than it feels at eleven at night with five screenshots open.

#Thirteen links point at me and I wrote all thirteen

The machine-readable identity on this site declares thirteen profiles — the site itself, GitHub, Hugging Face, LinkedIn, three syndication accounts, an ORCID, a speaker profile, a portfolio directory and the rest. Each one is declared three ways over. It is, technically, a complete job.

Every engine that ranked somebody else first did it off a page that person did not write. A national laboratory's staff directory. Forbes. A conference site. A university CV in PDF. Google Scholar. A radio programme. An author page on a bookshop.

Thirteen against that, and not one of the thirteen is a third party writing about me. You cannot corroborate yourself, and a fourteenth profile would not change the count of independent sources, which is zero.

#Three of those thirteen sit on domains a namesake already outranks me on

Then a search-index spot check the same week made it worse, and this is the part I did not expect.

Four cold queries, thirty-five result slots, this domain absent from all of them. Two directory sites returned on four queries out of four — both for somebody else with my name. And when I looked properly, three of the thirteen profiles I had so carefully declared are on platforms where a namesake already holds the top slot for our shared name. On one of them his account has years on mine.

A profile on a domain that somebody else already owns the top result for is not an independent source corroborating you. It is a thinner page filed underneath an established one on the same host. It is still worth filling in — an empty profile is worse — but it should be counted for what it is, and I had been counting it as progress.

So the honest state of it: the on-site half works, and ChatGPT citing four of my pages by title is the proof that it works. The half that decides everything else is other people writing about me, of which there is currently none, and which is the one part of this I cannot do by editing a file.