When an AI assistant answers a question about your market, its answer stands on pages. Some are yours, most are not. The pages you write are your surfaces: your site, your listings, your profiles, anything where you hold the pen. The pages where someone else writes about you are corroborations. Both matter, and they are won in completely different ways.
This guide is about the second half. It draws on 923 AI answers we archived in July 2026 across four engines, with every source they cited.
On trust questions, third parties answer for you
When we put identity and trust questions about a brand to the engines ("can I trust them", "what do people say about them"), the citations do not go to the brand's own site. They go to review platforms, professional profiles, third-party registries and verification pages. The detail per engine is in our study What AI assistants actually cite.
That result meets another one, measured on other markets: the sources that carry a verdict are often modest. Across our 923 answers, the engines cited 2,049 different root domains, and the ten most-cited domains of an engine carry only 12% to 43% of its citations depending on the engine. The set of pages an AI answer can stand on is wide, and small sites are inside it.
Where a corroboration pays the most
Two facts from the same corpus tell you where to aim.
On buying questions, the engines lean on lists and comparisons. On the engines that expose the page they read, 41% of the citations we collected on buying questions ("which tool should I use for X") point at a comparison, a ranking or a category listing published by somebody else. If your market has those pages, being in them is the highest-yield corroboration available.
Each engine reads a different web. In our measurements, ChatGPT leans toward institutional sources and established media; Gemini spreads across specialist operators, blogs and video; Perplexity reads platforms first; Claude stands on pages signed by an expert or a practice. The same corroboration is therefore not worth the same everywhere, and that reads engine by engine, never on average.
The families of sources you can obtain
A corroboration is obtainable when the gesture is bounded and known: a submission, a request, a proposal, a membership. That criterion, not prestige, is what separates a reachable source from a wish. The families below are portable from one market to the next; the specific sites change with the industry and the language.
- Open lists maintained by a third party: curated lists, reference repositories, sector wikis. You propose, the maintainer publishes under their name.
- Directories, comparison sites and review platforms: the record is a surface, the reviews and the comparison pages derived from them are corroborations.
- Interviews, podcasts and niche newsletters: many are looking for guests. These are doors of a person more than of a brand, so they open to the founder or the in-house expert.
- Contributive knowledge bases: an entity documented with its sources, reviewed and corrected by third parties.
- Editorialised sector and local directories: their editorial team writes the entry.
- Registries and semi-official databases: declarative, often automatic, and read when the question is about existence and legitimacy.
- Third-party video that cites you: what the engines read there is the transcript, as our study YouTube in AI answers shows.
- Partner, integration and customer-story pages at a supplier you already use: the third party has its own reason to publish.
The line that decides: who holds the pen
The canonical example is the review platform. The record is a surface, since you write it; the reviews are corroborations, since your customers write them. Paying for the record changes nothing about that border: the editing hand decides.
Two practical consequences. Posting in a community yourself is not a corroboration, it is one more surface placed elsewhere, and neither is a press release republished as is. Conversely, a page that mentions you without a link is still a corroboration: the mention and the link are not the same lever, and our study Backlinks and AI visibility measures what each one carries.
Choosing: listening on one axis, obtainability on the other
A target is judged on two axes. Listening: does any engine read this source on your market's questions? Atlas ranks the sources of your answers by AI Authority, so by measured listening. Obtainability: what does the gesture cost, and is it bounded? A source with moderate listening where the placement can be had today often beats a prestigious one out of reach, because the denominator is effort.
What does not return that effort: bought links, site networks, directories with no editorial team, blog comments and serial guest posts. Those sources are barely read by the engines, and the gesture turns against you once it is spotted.
Making what you obtain count
A mention obtained and never recorded disappears from your steering. The gesture has three steps, and all three are tooled.
- Record the corroboration in the Corroborations registry, dated by its publication rather than by the day you found it: it is the publication date that compares to the curves.
- Date the action in the Logbook. It appears as an annotation in front of your citation curves, so what moved reads against what you did.
- Measure again with Tracking, on the same questions, at the same rhythm. A series that continues through your actions is what makes a change readable.
All of it is driven from the app, the REST API or the MCP server, so your own AI assistant can keep the registry for you.
Method
923 answers collected between 2026-07-10 and 2026-07-25 by Epovest trackers through the official APIs of ChatGPT, Claude, Gemini and Perplexity, on fixed question panels phrased as customers ask them. Every answer is archived with its cited sources; domains are normalised to their registrable root. The share of comparison and listing pages is computed on the engines that expose the path of the page they read. Figures may be reused with attribution.