How AI answers pick brands

Nobody outside an AI engine can see how it picks the brands it names, so the useful question is what you can observe. An AI answer is one composed response to a question. It may name brands, give each one a role such as top recommendation or warning, and link a few pages as its sources, sometimes after running searches of its own. Appearing (being named), being cited (your domain linked as a source) and being recommended (the role you were given) are three separate results, each counted against its own total. What else you can see, such as the searches behind an answer, depends on how it was collected: through a public app, an engine's API, or Google's AI surfaces.

In short

• An AI answer has parts you can record: the answer text, the brands it names, the role it gives each brand, the sources it cites and, on some engines, the searches it ran first. None of them tells you why the engine wrote what it did.
• SEO still matters, because authority, structure and freshness carry over. But ranking a page and being quoted inside an answer are different achievements, and one does not guarantee the other.
• A mention, a citation and a recommendation are separate observations with different denominators. An answer can give you any one of them without the others.
• Apps, APIs and Google's AI surfaces are different collection channels. They receive your prompt differently, expose different evidence and can answer the same prompt differently, so report each one on its own.
• Start by measuring a few things well: per engine, brand visibility, citation rate and the recommendation roles behind your mentions, each with its count and denominator, plus the shown rate on Google's surfaces.

What changed when search started answering?

A search engine returns a ranked list of links for the reader to scan. An answer engine writes one response and names a handful of brands and sources inside it, so the unit of visibility moved from a ranked link to a mention or a citation inside the engine's text.

That shift changes what losing looks like. With a list, a reader can scan past the first result to a lower one. With a composed answer, the reader often never sees a list at all (GEO versus SEO), so if your brand is not in the answer, there is no second-place result to fall back on.

The names for this work overlap without being the same. Answer engine optimization (AEO) is shaping content so it becomes the direct answer to a question, and it aims at a single slot, most often a featured snippet, a voice response or a knowledge panel. GEO aims at being mentioned or cited inside the longer answers a generative engine composes from several sources. Both rest on the same foundations: clear, authoritative content and consistent entity information (What is AEO?). This guide covers AEO and, where the two overlap, GEO: getting named, cited and recommended in AI answers.

The thing you measure is a prompt: a full question you chose to monitor because your buyers ask something like it. A prompt is not a keyword, and it is not something a particular user typed (Key terms). Choosing them is the subject of Choose the prompts to track.

Working with answers needs names for their parts. Figure 1.1 labels the parts of one answer that can be observed and recorded.

• Answer text. What the engine wrote. What counts as the text differs by engine: the stored ChatGPT (app) answer includes the product lists and local business lists the page shows, while product cards, ads and videos inside a Google AI Overview are left out of its text (What counts as the answer).
• Brand mention. Your brand named in the answer. A mention says nothing yet about whether the naming was favourable.
• Recommendation. The role the answer gave a brand once it named it, such as Top recommendation, Alternative option, Warning or caveat, or Not recommended. A role is independent of sentiment, and a brand named as a warning is still a mention (Recommendation roles).
• Cited source. One link the engine puts in its own answer, pointing at a specific URL as the source for what it just said (Citations). Not every link qualifies. A link on a company name in a list, a product card and a map place are different observations, and A source link, a brand link, and a shopping card are different observations shows how to code each one.
• Fan-out search. Some engines expand one prompt into several searches of their own before they answer. A monitoring tool can capture that expansion only where the engine exposes it.

One more idea sits outside the visible answer. A retrieval is a page the engine pulled into its context without linking it in the answer (Key terms). You can only see retrievals on engines that report the pages their search or grounding brought back.

Every one of these parts is an observation of what the answer contained. None of them is the reason it contained it. An engine that names a competitor first has not told you why, and a report that says "AI ranks us lower because" claims more than the evidence holds.

What does SEO still do?

SEO still does the groundwork. Authority, clear structure and fresh content carry over to AI answers, and on Google's AI Overviews the organic results sit on the same page as the answer. What SEO does not do is guarantee a citation, because ranking a page and being quoted inside an answer are different qualities.

The two disciplines measure different things. SEO aims to rank a page high in a results list and reads rankings, impressions and clicks. Answer visibility aims to be mentioned or cited inside a composed answer and reads mentions, citations and share of voice (GEO versus SEO). A rank tells you where a link sat. It does not tell you whether an answer named you, credited your page or recommended you.

Some SEO habits mislead. Chasing a keyword's exact phrasing does little for an engine that paraphrases the question before it answers, and optimising purely for position ignores that a page can rank first and still never be quoted.

That gap is the one to look for. A brand can hold a top organic position for a query and still not appear when ChatGPT or Perplexity answers the same question, and the gap stays invisible unless both are measured side by side (Google rankings versus AI citations). The docs treat proven authority in Google paired with absence from the answer as a fast opportunity rather than a cold start: the page does not need to build authority from nothing, it may need its argument restated in a form an engine can quote. Treat any such change as a hypothesis to test; Decide what to fix first covers how to choose and check one.

Google's own AI Overviews let you set the two side by side directly. Every Google AI Overviews run in DiscoveredBy keeps the first page of organic results for the same search, whether or not an Overview was shown, so you can read organic position next to whether the Overview cited you. Google AI Mode returns no organic results, so this comparison covers AI Overviews only (Engines and measurement).

Your search data is still a starting point, but a keyword is not a buyer question; Get your brand ready shows how to turn one into a prompt.

Appearing, being cited and being recommended: why are they different?

They are three different observations of the same answer, counted from different populations. A mention is your brand named, a citation is your domain linked as a source, and a recommendation is the role the answer gave you once it named you, so an answer can be strong on one and empty on another.

The populations matter as much as the events. A collected answer is a completed prompt execution: the engine ran and returned a result. A run where Google showed no AI answer is not a collected answer. An analysed answer is a collected answer whose brand mentions have also been extracted, so it is always a subset of the collected group (Collected versus analysed answers).

Each idea then has its own metric:

• Brand visibility is analysed answers naming your brand, divided by analysed answers (Brand visibility).
• Citation rate is collected answers that cited your domain, divided by collected answers (Citation rate).
• Domain coverage, the retrieval measure, is collected answers that retrieved your domain, divided by collected answers. It has no value on engines that report only the sources their answer cites.
• Recommendation roles are counts of mention records by role. Any rate you build from them is your own calculation, so state its denominator.

Mentions and citations combine in four ways, and Figure 1.2 sets out what each combination means.

Named and cited. The answer named you and linked one of your pages. It is the strongest of the four on paper, but the link does not show that your page shaped what was said, and you should read the role before celebrating: brand visibility counts a mention as a warning the same as a mention as the top recommendation.

Named, not cited. The answer named you but linked other pages, or none. That is real visibility. What it cannot tell you is where the engine's description of you came from, which is the question Your brand is mentioned without a source works through.

Cited, not named. The answer linked one of your pages as a source, yet your brand is not named. An answer might explain a review workflow, link your guide and then recommend three other tools. Your website is cited, but your brand is missing from the recommendation covers that case. There is also a counting rule to know here. On ChatGPT (app) and Gemini (app), a small source link whose text is only a site address is blanked out before mentions are extracted when it directly follows the passage it supports, so a site named only in such links is not counted as a mention, although the stored answer keeps the link (Engines and measurement).

Neither. You are absent from this answer. Before you read that as a loss, check there was an answer to be in: a Google surface may have shown no AI answer, a run may have failed, and an engine may have cited no sources at all. Run your first audit works through how to tell those apart.

Recommendation sits on top of the two combinations where you were named. A mention is not a good outcome by default, so report mentions next to roles, never instead of them.

Illustrative example: Quillstone and its competitors are fictional, and the numbers are made up to show the method.

Quillstone sells document-review software to mid-sized legal and compliance teams and tracks three competitors: Brieflane, Clausewise and Docket North. On one engine that reports the pages it retrieved, over one 28-day window, it has 60 collected answers, of which 50 have been analysed.

• 20 of the 50 analysed answers name Quillstone, so brand visibility is 40%.
• 9 of the 60 collected answers cite a Quillstone page, so citation rate is 15%.
• 14 of the 60 collected answers retrieved a Quillstone page, so domain coverage is 23.3%.

Both 40% and 15% are correct at the same time: they count different events and divide by different populations. If the 9 citing answers are among the 14 retrieving ones, 5 answers pulled in a Quillstone page and did not link it, which is a question in itself; Your page was retrieved but not cited takes it from there.

The 20 mention records split into 6 Top recommendation, 9 Alternative option, 3 Warning or caveat, 1 Not recommended and 1 Unclassified, which adds back to 20. So 4 of the 20 mentions are a warning or a "not recommended". Six top recommendations is 30% of the 20 answers that named Quillstone and 12% of the 50 analysed answers. Both figures are true, they answer different questions, and neither is a built-in metric, so a report that uses one should name its denominator and label it a hand calculation.

Share of voice behaves differently again. It divides your brand's count by the summed counts of every active tracked brand, yours included, so adding or pausing a competitor moves it with no change in the answers (Share of voice). AI visibility, brand mentions, and citations: what each number tells you works the same Quillstone dataset through every metric, share of voice included.

Why do apps, APIs and Google's AI surfaces give different answers?

Because they are different collection channels. Each reaches the engine by a different route, receives your prompt with different context and returns different evidence, so the same prompt can produce different answers, and each channel is its own population of answers.

A collection channel is one provider, platform, surface and collection method taken together. DiscoveredBy's eight engines fall into three groups by how they are collected, documented in its engines reference; other tools may collect differently. Figure 1.3 sets the three channels side by side: how each is collected, what each answer records, and what it cannot show.

The app channel. ChatGPT (app) and Gemini (app) are collected by DataForSEO, a third-party data provider, which puts the prompt to chatgpt.com or gemini.google.com in a session that is not signed in and returns the answer the page shows. The product labels this collection method "Licensed data". The prompt arrives as written, on one line, with nothing added: no audience block and no "Search context" block, so persona variants do not run. The country travels as a setting of the request. Gemini (app) runs a city target as the city's coordinates; ChatGPT (app) has no city setting, so city targets do not run on it. Each app decides for itself whether to search, and DiscoveredBy never asks it to, so some answers cite no sources at all. What these answers record: the page's text, the sources it shows as citations, and answer features such as shopping and local businesses. ChatGPT (app) also reports its searches and the pages they found when it chooses to search, from 2026-09-29. What they cannot show: what a signed-in person sees, the searches behind a Gemini (app) answer, or a comparison of retrieved and cited pages: Gemini (app) lists no pages searched, and the pages ChatGPT (app) found are stored but not yet compared (Retrieved vs cited).

The API channel. Gemini (API), Perplexity, Claude and Grok are queried through their own company's API. The message carries the prompt plus any audience, response-language and "Search context" blocks, so persona and language variants run on them. The evidence differs by engine. Claude reports the results its search tool returned. Perplexity returns every source it retrieved and marks the ones its answer used, but not the text of its searches. Grok reports the pages it opened and the sources it listed. Gemini (API) is called with Google Search grounding and reports the pages it grounded its answer on. What an API answer cannot show is what the company's own app shows a person. Gemini (API) and Gemini (app) are separate engines whose answers to the same prompt can differ, and in every one of DiscoveredBy's checks on 2026-09-28 the data provider reported a model for the app that is not the one Gemini (API) calls. Investigate a disagreement between Gemini app and API answers shows how to work through a gap between the two.

Google's AI surfaces. Google AI Overviews and Google AI Mode are also collected through DataForSEO, which fetches Google's own result: the results page with the AI Overview and its organic results, or the AI Mode answer, which has no organic results. Google receives only the search text, so persona variants do not run, and neither do Chinese variants or prompts over 700 characters once encoded; each is left out rather than sent in a changed form. Google does not say which model wrote the answer. These surfaces have one outcome the others lack: when the request succeeds but Google shows no AI answer, the run is stored as no AI answer shown, left out of every answer-based number and reported instead as the shown rate, the share of runs on which Google showed an AI answer (When Google shows no AI answer). What they cannot show: the searches behind the answer, which read as not applicable, or a list of pages searched. AI Mode also reports no answer feature the product counts. Google AI Overviews, AI Mode and Gemini: plan separate measurements gives each of the four Google-owned engines its own measurement plan.

Location is a fourth source of difference. A target location is information passed on, and each engine decides how much weight to give it, so the answer can differ from what a real person there sees, signed in, on their own device (A target location is not a customer's location).

Two habits follow from all this.

Report each channel separately. Pooling two channels sums their numerators and denominators into a blend that neither channel produced. A pooled number can fall when a channel with a lower rate joins the window, while neither channel's own rate moved, and it can stay flat while one channel rises and the other falls. Report one figure per channel, and record any channel added, removed or relabelled in a dated note. Keep app results and API results separate in your reports gives a convention you can copy, and Keep a measurement change log when an AI engine changes covers the log.

Treat your own check as one more channel. When you type a question into ChatGPT yourself, you are probably signed in, on your own device, in your own location, with wording you chose on the spot, at one moment. A monitoring result is a fixed question asked on a schedule under recorded conditions. Before deciding which is wrong, log seven variables for both: session, wording, timing, location, language, collection method and what counted as the answer. To compare fairly, change your check to match the tool, by pasting the exact prompt text into a signed-out session, and call the result a closer match, not a reproduction. Why your manual ChatGPT check differs from a monitoring result has the discrepancy log.

What can you learn from an answer with no web search?

You can still learn what the answer said: which brands it named, in which roles, and which sources it cited, if any. What a blank search record cannot tell you is where the answer's content came from, and the blank itself can mean three different things.

Some engines decide for themselves whether to search. ChatGPT (app) and Gemini (app) are never told to, so some of their answers cite no sources, and that is normal behaviour rather than a fault. Other engines never expose their searches. And sometimes the capture simply did not run. DiscoveredBy derives a fan-out state for every answer from the queries and sources on hand, and four of its seven states are where blanks live (Fan-out). They sort into three cases.

The engine answered without searching. Capture ran and found no queries and no sources, so the state reads no search. This is the one blank that is a claim about the engine, and its scope is that run. An engine that chooses whether to search may search on the next run of the same prompt.

The engine does not report its searches. The state reads not applicable for Google AI Overviews, Google AI Mode and Gemini (app), which report only the sources their answers cite, and for Perplexity answers collected through its earlier Search API. That blank is a statement about the surface, not about the answer.

The record has a gap. Searches not recorded means sources exist for the answer but no queries do. Perplexity's Sonar collection returns the sources it found but not the text of its searches, so any Perplexity answer with a source reads this way, and a returned source counts as evidence that a search happened. Not recorded means fan-out capture did not run for that answer at all; ChatGPT (app) answers collected before 2026-09-29 read this way.

None of the three tells you the answer came from training data, memory or any other origin. Even an observed no-search only says that on this run the engine reported no searches and returned no sources. Report it in those words, and add one line to any report that uses them: no recorded search does not show where an answer's content came from.

What remains is still worth reading:

• The text and its mentions. Whether you were named, in what role and with which stated reasons is observable whatever the search state.
• The cited sources, if any. On ChatGPT (app), a cited page is always among the pages found.
• The channel. Which collection channel produced the answer decides which of the cases above can apply.
• Repetition. Because an engine may search on one run and not the next, the same prompt across several days is the honest way to see a pattern.

This also changes how you read a citation rate. An app answer that cited nothing could not have cited you, so read your citation rate beside answers citing sources: collected answers that cited at least one source of any domain, divided by collected answers. It describes the engine, not you. It has no value for Claude and Grok, whose saved citations include every search result (Answers citing sources). The answer shows no web search: how to report what evidence remains has a decision tree and reporting phrases for each state.

What should you measure first?

Measure a few things well before you measure many. For a small, fixed set of prompts, read each engine on its own and record brand visibility, citation rate and the recommendation roles behind your mentions, each with its count and denominator. On Google's AI surfaces, read the shown rate first.

A first pass, in order:

1. Write the engine coverage note. Before the first report, write down which engines you track, how each is collected, what each answer records and what it cannot show. The worksheet below is a template.
2. Read one answer per engine by hand. A dashboard summarises answers you have not seen. Run your first audit starts with a single answer for that reason.
3. Record the populations. For each engine and window, note the collected and analysed answer counts. Every rate after this divides by one of them.
4. Read three numbers per engine. Brand visibility over analysed answers, citation rate over collected answers, and the role counts behind your mentions. Put answers citing sources beside the citation rate as context.
5. Read the shown rate on Google's surfaces. A low citation rate on Google AI Overviews can mean you are rarely cited, or that Google rarely shows an Overview for your prompts.

Continuing the Quillstone example, take one prompt on Google AI Overviews over 20 daily runs. An Overview appeared on 12 of them, a shown rate of 60%. Quillstone's domain was cited in 5 of those 12 Overviews, which is 42%. Dividing the 5 by all 20 runs would give 25%, and would blame Quillstone for Google's decision not to show an Overview on the other 8 runs. Quillstone's domain also sits on the first page of organic results on 9 of the 20 runs, or 45%, counting runs with no Overview. Ranking on 9 runs and being cited in 5 of 12 Overviews is a hypothesis to test, for example by opening the runs where it ranked but was not cited. It is not a conclusion about why. Twelve answers is also a small sample, so 42% is a reading to watch rather than a settled figure.

A few rules keep these first numbers honest:

• Put the count beside every percentage. "5 of 12" says more than "42%".
• Treat small samples as provisional. The docs mark a metric built from fewer than 30 observations as provisional: still shown, but too small to read as settled (How to read any number).
• Report no data as no data. A rate with no denominator reads as no data, never as 0%, and a blank cell for something an engine does not report is not a zero either.
• Describe observations, not causes. "Named in 20 of 50 analysed answers" is a finding. "The engine prefers Brieflane" is not.

Some measures can wait. Share of voice depends on the competitor list you set, which Get your brand ready helps you choose. Trends and before-and-after comparisons need a fixed prompt set and a change log, which Measure change honestly covers, along with Use a fixed prompt cohort for month-to-month comparisons.

Engine coverage note

Write this note once per project, keep it at the top of your reporting template, and add a dated line whenever an engine or a collection method changes. The example blocks are filled from DiscoveredBy's engines reference as it stood on 2026-10-01; check that page for the current state, or replace the blocks with what your own tool documents.

ENGINE COVERAGE NOTE
Project: ________   Prompt set: ________   Countries / cities: ________
Written by: ________   Date: ________   Engines reference checked on: ________

Fill one block per engine you track. The first four lines describe the
channel; the last four are your decisions.

ENGINE: ________
  How collected:       own API call | licensed data provider (app or search page) | other: ____
  What reaches it:     prompt text only | prompt + audience / language / location text
                       location: as a setting | in the prompt | not sent
  What each answer records (tick):
                       [ ] answer text            [ ] brand mentions and roles
                       [ ] cited sources          [ ] searches (fan-out)
                       [ ] returned or retrieved pages
                       [ ] organic results        [ ] answer features
                       [ ] "no AI answer shown" outcome
  What it cannot show: ________
  Variants that do not run here (persona, language, city, long prompts): ________
  Primary metric and its denominator: ________
  Metric we will NOT read on this engine, and why: ________
  Compared only within this engine: ________

EXAMPLE BLOCKS (from DiscoveredBy's engines reference, 2026-10-01)

ChatGPT (app)
  How collected:  licensed data provider; chatgpt.com, session not signed in
  What reaches it: prompt text only, one line; country as a setting; no city
  Records:        page text incl. product and local business lists; inline
                  sources as citations; searches when it searched (from
                  2026-09-29); answer features
  Cannot show:    a signed-in person's answer; a retrieved vs cited rate
                  (cited sources only); city, persona and Chinese variants,
                  and prompts over 2,000 characters (not run)
  Note:           decides for itself whether to search; an unreadable
                  answer is a failed run, never "no AI answer shown"

Gemini (app)
  How collected:  licensed data provider; gemini.google.com, not signed in
  What reaches it: prompt text only, one line; country as a setting; city
                  as coordinates
  Records:        page text; web pages among its sources as citations;
                  Google product and Maps place links as shopping and
                  local business features
  Cannot show:    its searches (fan-out not applicable); a list of pages
                  searched; persona and Chinese variants, and prompts over
                  2,000 characters (not run)

Google AI Overviews
  How collected:  licensed data provider; Google's results page
  What reaches it: search text only; country, or city coordinates, as a setting
  Records:        Overview text and cited sources; first page of organic
                  results on every run; "no AI answer shown" (shown rate)
  Cannot show:    its searches (not applicable); pages searched; which
                  model wrote it; persona and Chinese variants, and prompts
                  over 700 characters (not run)

Google AI Mode
  How collected:  licensed data provider; Google AI Mode's answer
  What reaches it: search text only; country, or city coordinates, as a setting
  Records:        answer text and cited sources; "no AI answer shown"
  Cannot show:    organic results; answer features; its searches; pages
                  searched; persona and Chinese variants, and prompts
                  over 700 characters (not run)

Gemini (API)
  How collected:  own API call with Google Search grounding
  What reaches it: prompt plus audience, language and "Search context"
                  blocks; location in the prompt text only
  Records:        grounding sources, credited to the domain Google's
                  redirect lands on; the searches Google reports
  Cannot show:    the answer gemini.google.com shows (they can differ);
                  a retrieved vs cited rate (grounding sources only, not
                  compared); a citation position comparable with Gemini (app)

Perplexity
  How collected:  own API call (Sonar)
  What reaches it: prompt plus audience, language and "Search context"
                  blocks; location also as a setting
  Records:        sources its answer marks, as citations; every returned
                  search result
  Cannot show:    the text of its searches ("searches not recorded");
                  what Perplexity's own app shows

Claude
  How collected:  own API call with its web search tool
  What reaches it: prompt plus audience, language and "Search context"
                  blocks; approximate location on the search tool
  Records:        its searches; every result its search tool returned,
                  each also stored as a citation
  Cannot show:    answers citing sources (no value); what Claude's own
                  app shows

Grok
  How collected:  own API call with web search
  What reaches it: prompt plus audience, language and "Search context"
                  blocks; location in the search tool's fields
  Records:        its searches; the pages it opened and the sources it listed
  Cannot show:    its individual search results; answers citing sources
                  (no value); what Grok's own app shows

CHANGE LOG
Date       | Engine        | What changed (added, removed, relabelled, method) | Reports affected
__________ | _____________ | ______________________________________________ | ________________

In DiscoveredBy

The Engines and measurement reference lists the eight engines DiscoveredBy measures, says whether each is queried through its company's API or collected as licensed data, and sets out what reaches each engine and what counts as its answer; it is the source for the example blocks in the coverage note. The Citations screen shows which pages the answers credit. It opens on the share of the citations in view that point at your domain, can be narrowed by engine, tag, country, persona and language, and lists the same citations grouped by URL or one row per citation. That share divides by citations, not by collected answers, so it is not meant to match your citation rate.