Search used to return links and leave the conclusion to the reader. Now it draws the conclusion. Type a name into Google and, more often each month, an AI Overview answers above the results; switch to AI Mode and the whole page is an answer. Ask ChatGPT, Gemini or Perplexity and there are no links at all until you ask for them. For a person, that means the first impression is no longer a list you can influence at the edges. It is a paragraph, written by a machine, from whatever it could find and trust.
The numbers behind this are large. A clickstream study published in 2026 found that 68% of US Google searches ended without a click, up from 60% two years earlier, that AI Overviews appear on more than a fifth of searches and cut click through by nearly 60% when they do, and that AI Mode had passed a billion monthly users. Whatever that answer says about you is, for most of the people looking, the whole result.
Two kinds of machine
It helps to separate the systems, because they fail differently.
Assistants that search the web live. Perplexity, Google's AI Overviews and AI Mode, Gemini with search, ChatGPT with browsing on. These run a search, read the top results, and write a summary with citations. They reflect the web as it is this week. Fix the source and they change within weeks of the fix being indexed.
Assistants answering from training data. ChatGPT without browsing, Claude without web access, and any model when it decides it does not need to search. These answer from what was online when the model was trained, which can be a year or more ago. Fix the source and they change when the model is refreshed, which you cannot schedule.
The practical consequence: the same question gets different answers from different tools, and a thin or wrong answer from one does not mean your record is failing. The fifteen question test is designed to show which kind of failure you have.
How a live answer gets built
- The name is resolved to an entity, if one exists. Google checks its Knowledge Graph for a confident match. If there is one, the answer is anchored to it and the panel data comes along. If there is not, the system falls back to reading pages and guessing, which is where namesakes win.
- A handful of pages are retrieved. Typically the top results for the name, plus pages that mention the name with context. Your home page, if it exists and is indexed, is one of them. A directory, an old conference bio and a stranger's LinkedIn are candidates too.
- Facts are extracted and weighed. Pages that state facts plainly, in sentences and in structured data, are easier to extract from than pages that imply them. Pages that agree with each other are weighted up. A single page that disagrees with the rest is usually ignored, unless it is the only one with the fact.
- The answer is written, with citations. The system picks two to five sources to show. These are the pages it leaned on, and they are the ones a reader will click. If your own site is not among them, the answer was written about you from someone else's description.
Why a namesake wins
Because the machine is resolving a string to a thing, and the thing with more corroboration wins the string. A dentist with your name who has a practice site, a healthgrades profile, two hundred reviews and a local news mention is a better documented entity than a consultant with a LinkedIn and a Canva one sheet. The fix is not to out shout the dentist. It is to become unambiguous: your facts on your own domain, a Person node that says which one you are, profiles that agree, and third party pages that repeat it. Disambiguating details help more than they seem: city, company, middle initial if you use it, the field you work in, stated in the first sentence.
What gets cited
Looking across the assistants that show sources, the pages that get cited for a person are consistent: a page on the person's own domain that reads as the canonical description, a Wikipedia or Wikidata entry where one exists, a recent interview or profile on a named site, and occasionally a LinkedIn. What does not get cited: link in bio pages, PDF one sheets, social posts, and anything the system cannot read because it is behind a login or rendered by a script.
This is why the record has to be on your own domain, in plain text, with the machine layer underneath it. A beautifully designed site with no structured data and a bio that implies rather than states will be read, but it will not be the page the machine chooses to lean on.
Four tools, four behaviors
The assistants are not interchangeable, and knowing how each behaves saves you from misreading the results.
Google AI Overviews and AI Mode. Anchored on the Knowledge Graph when an entity exists, otherwise synthesized from top results. Most sensitive to structured data and to Wikidata, because those feed the graph directly. Slowest to show a new entity, fastest to reflect a change once the entity exists.
Perplexity. Pure live retrieval with visible citations. The most transparent of the four: it shows you exactly which pages it leaned on. If your own site is not cited, it was not among the top retrievable pages for your name, which is a ranking and indexing problem before it is anything else.
ChatGPT. Two modes that give different answers. With search on, it behaves like Perplexity with less visible sourcing. With search off, it answers from training data and will confidently repeat whatever was online at the last training cut, which can be a year or more stale. Ask it which mode it used if you are unsure.
Gemini. Blends Google's graph with live search and its own model knowledge. Tends to give the panel's facts when a panel exists and to hedge when one does not. The best early indicator that Google has formed an entity for you, often before the panel shows on a plain name search.
Claude. Answers from training data unless web access is enabled, and is the most likely to say plainly that it does not have reliable information, which is a more useful answer than a confident wrong one. Treat its answer as a measure of what the web said at its training cut.
What "cited" means and why it matters
When an answer shows sources, those pages did two things: they were retrieved for your name, and they contained facts the model could extract. A page gets retrieved by being indexed and ranking for the name. It gets extracted from by stating facts plainly: a sentence that says where you were born beats a timeline graphic that implies it. The pages that win both are canonical personal sites with prose and structured data, Wikidata, and recent interviews on sites with clean HTML.
A citation is also a click. Someone reading an AI answer about you and clicking through lands on the cited page. If that is your site, you control the next impression; if it is a directory, you do not. Being cited is the modern equivalent of ranking first, and it is earned by being the most extractable source about yourself.
What a person can do, in order
- Run the test and keep the screenshots. Fifteen answers, dated. You cannot improve what you have not measured.
- Build the canonical page. Facts in sentences. One Person node. A sameAs list. Disambiguation in the first paragraph.
- Make the profiles agree. Same name, same photo, website field pointing at your domain.
- Add corroboration that gets cited. Interviews, a profile piece, a press release, listed on your record with links. These are the pages live systems lean on.
- Request indexing and wait. Search Console, then weeks for the live systems and longer for the trained ones. Run the test again at ninety days.