Definition
What Is AI-Mediated Search?
AI-mediated search
AI-mediated search is search in which an AI system interprets the request, gathers sources, and composes an answer before the person reaches a website. Instead of working through a page of ranked links, the person is presented with a composed answer, a limited set of named options, and the sources behind them.
Traditional search returned a page of links and left the comparison to the customer.
AI-mediated search performs more of the interpretation, research, and comparison first, then presents a shorter set of options.
The Surfaces
Where does AI-mediated search happen?
AI-mediated search happens in two places: inside Google Search, and inside assistant products people use as a search entry point.
Inside Google Search
AI Overviews and AI Mode sit inside Google Search itself rather than in a separate product. Google Search Central documents AI Overviews as helping people get to the gist of a complicated topic or question more quickly while providing a jumping off point to explore links, and AI Mode as particularly helpful for queries where further exploration, reasoning, or complex comparisons are needed.
- AI Overviews
- AI Mode
Assistant products used as a search entry point
The same interpretation happens inside assistant products rather than a results page. ChatGPT, Perplexity, Gemini, and Microsoft Copilot are separate products from separate companies. OpenAI and Perplexity each publish documentation describing the crawler that surfaces websites in their search results and how a site controls its access.
- ChatGPT
- Perplexity
- Gemini
- Microsoft Copilot
| Surface | Where it sits | Observed by Hendricks |
|---|---|---|
| AI Overviews | Inside Google Search | Yes |
| AI Mode | Inside Google Search | No |
| ChatGPT | Assistant product | Yes |
| Perplexity | Assistant product | Yes |
| Gemini | Assistant product | Yes |
| Microsoft Copilot | Assistant product | No |
Which surfaces does Hendricks observe?
Hendricks observes four systems: Google AI Overviews, ChatGPT, Perplexity, and Gemini.
Google AI Mode and Microsoft Copilot are surfaces that exist in the same information environment, and they are named here for that reason alone. Hendricks does not measure, test, monitor, or report on Google AI Mode or Microsoft Copilot.
Ranking Versus Appearing
Why can a brand rank #1 on Google and still not appear in AI answers?
A brand can rank #1 on Google and still not appear in AI answers, because a position in a ranked list of links and a named source inside a composed answer are two different outcomes, and holding the first does not produce the second. Hendricks publishes no figure for how often this happens. The share of AI-cited pages that also rank in Google is a different quantity from the share of well-ranked pages that get cited, and only the second describes this problem. What can be established for one brand is narrower and more useful: whether it was named, on which observed surface, for which question, and on which date.
- The brand is absent from the options presented, so it is never evaluated.
- The brand appears, but is described from outdated, thin, or contradictory evidence.
- The brand is mentioned without being recommended, and the shortlist forms around competitors.
- The decision narrows before any website visit, so no analytics event records the loss.
- Rank and traffic reporting stays flat while consideration falls, because the two measure different things.
- Nobody in the business can say whether the brand was considered, because nothing in the stack observes the surfaces where the comparison happened.
Hendricks calls this the AI Selection Problem.
Being discovered and being chosen are now separate outcomes, and most reporting still measures only the first.
Absence Versus No Sources
Why is my brand not showing up in ChatGPT?
A brand is missing from a ChatGPT answer for one of two reasons that look identical on screen and are not the same problem. Either the answer cited sources and the brand was not among them, or the answer cited nothing at all. An answer that cites nobody has no slot for any brand to win.
Hendricks measured that split on its own brand and published the run. On 2026-08-19, all three engines were sent the same 17 buyer questions. ChatGPT cited sources on 2 of them and answered the rest with no source attached. Perplexity cited sources on 17. Google AI Overviews returned a measurement on 13 of its 17 and cited sources on 1 of those. The run produced 51 cells, measured 47 of them, and 20 of those carried a citation of any kind. It is archived as run 2026-08-19-110930.
A later archived run, 2026-08-20-110653, published on the study of whether any source was cited by all three engines, recorded ChatGPT citing a source on 2 of 17 questions. Those were the same two questions across six dated runs. That count is citation presence only.
Those figures describe 17 questions, three systems, one geography, and one date. Nothing was changed between that run and the one before it, and nothing was held back for comparison, so the run establishes what those answers looked like and nothing about what a change to a website would do to them. It is a baseline, not evidence that any tactic works.
What the split changes is the first question worth asking. Before a brand asks why it lost a ChatGPT answer, it has to establish whether the answer had anything to lose. For most of the questions in that run, the honest reading is not that a shortlist formed without the brand. It is that no shortlist was published at all.
An answer that names a source is not the same as an answer that is right. In the run a day earlier, ChatGPT answered a question about who audits brand visibility in AI assistants by listing viaudit.com, a domain that returned no DNS record and no response when it was checked with dig and curl on 2026-08-18.
Ruling Causes Out
How does a brand find out why it is not showing up in ChatGPT?
A brand finds out why it is not showing up in ChatGPT by separating four states that look the same on one answer screen, in order, cheapest first. Three of the four are not competitive losses. Each carries a different remedy, and only two of them respond to anything a brand does to its own evidence.
| Check | What it establishes | Why it sits here |
|---|---|---|
| Did the answer cite any source at all? | Whether the question produced citation slots that any brand could occupy. | An answer that cites nobody is not a shortlist the brand lost, and no work on the brand adds a slot to it. |
| Did a sourced answer leave the brand out? | Whether the brand was passed over inside an answer that named other sources. | This is the only one of the four states that is a competitive loss, and it is the one most treatments of this question assume without checking the first. |
| Did the answer name the brand and describe it wrongly? | Whether the problem is absence or description. | A brand described from outdated, thin, or contradictory evidence reads as missing to anyone scanning an answer for a reason to shortlist it, and the remedy is a different one. |
| Did the result hold when the question was asked again? | Whether the reading is an observation or a single screen. | A result that appears in one run and not the next has not been established, so nothing read once settles which of the three states above applies. |
Two of those four states are worth acting on, and neither is fixed by publishing more pages. Absence from a sourced answer and a wrong description are both conditions of the evidence an AI system can find and corroborate about a brand: whether the site can be reached and read, whether the organization, its services, and its people are defined consistently wherever they are described, and whether decision-stage content exists for the questions buyers actually ask. Search Presence Engineering is the work on those conditions.
Hendricks labels the reasoning behind that work for what it is. Improving those conditions is a plausible mechanism rather than a measured effect. No controlled test available to Hendricks shows that the work produces a citation, no page on this site claims one, and the two published runs tested nothing, because Hendricks changed nothing about the site between them and held nothing back for comparison.
Traditional Versus AI-Mediated
How is AI-mediated search different from traditional search results?
Traditional search results return a page of links and leave the comparison to the customer. AI-mediated search returns a composed answer, and much of the comparison is finished before the customer clicks anything.
| Dimension | Traditional search results | AI-mediated search |
|---|---|---|
| What the customer receives | A page of ranked links | A composed answer with a limited set of named options and cited sources |
| Where comparison happens | Across several websites, after the click | Largely before the click, inside the answer |
| What a brand competes for | A position on the results page | Inclusion in the set of options presented |
| What the brand supplies | The page the customer lands on | The evidence a system can find and corroborate about it |
| Stability of the result | Comparatively stable for the same query | Can vary with context, wording, location, platform, and time |
| What measurement reports | Rank, impressions, clicks, and sessions | Observed consideration rate and observed recommendation rate under stated conditions |
Neither replaces the other. A brand can rank well and still lose the shortlist, which is why the two are measured separately.
Vocabulary
What is AI search visibility?
AI search visibility is the presence of a brand in the answers AI systems compose and in the sources those answers cite. It is a presence measure: it records that the brand appeared, on a named surface, for a named question, on a given date.
Visibility is not selection. A brand can be present in an answer that goes on to recommend a competitor, and it can be present for questions no buyer with budget ever asks. Entering a consideration set that carries commercial value is a separate outcome, earned separately, which is why Hendricks reports observed consideration rate and observed recommendation rate rather than a count of appearances.
Any visibility Hendricks reports is bounded by the surfaces marked as observed in the table above.
Generative engine optimization (GEO) and answer engine optimization (AEO) are names for the work aimed at earning that visibility. AI-mediated search is the environment that work is aimed at, not the work itself.
Everything Hendricks reports about these surfaces is an observation under stated conditions.
Honest limitation
- These surfaces change without notice, so a result observed on one date may not reproduce on the next.
- Output varies with user context, which means a single observation is not a ranking.
- Hendricks cannot see inside these systems. Hendricks reports what was observed, when, on which surface, and under which conditions, and makes no claim about how any system decides.
Sources
Sources
Last reviewed .
This definition is maintained by Hendricks. Where the page describes systems Hendricks does not control, it states publicly observable behavior and cites the platform’s own documentation. The figures on this page come from first-party Hendricks measurement, published in full with its query set, its denominators, its method, and its limits at the Hendricks Selection Baseline. No third-party research, vendor study, or statistic is reported.
This definition is applied in the Selection Intelligence solution and the measurement methodology.
Change History
Change history
| Date | Kind | Summary |
|---|---|---|
| Not yet recorded | Publication | First publication of this page. |