Cited but not recommended: why AI uses your content and names other brands

When an AI answer links one of your pages but recommends other brands, the link and the recommendation are two separate findings, and you investigate them separately. The citation shows your page was recorded as a source for something the answer said. The recommendation shows which brands the answer put forward. Compare the claim your page supports, the prompt's intent, whether your brand appears in the cited page's text, and how the answer framed each brand. What you find describes one answer. It does not show why the engine chose as it did.

In short

• A citation is a link to your page. A recommendation is a role the answer gives a brand. You can have the first without the second, and DiscoveredBy's docs treat them as different records.
• Start by reading what claim your cited page supports in the answer. An informational page can support an informational claim without supporting a product choice.
• Check the boring causes first: prompt intent, brand naming, and whether your name appears in the page's text.
• Compare against a small set of similar answers before you conclude anything from a single one.
• Write every finding at the strength the evidence supports. The sheet below has a column for that.

What does "cited but not recommended" actually mean?

It is a gap between two records of the same answer: your domain appears among the sources, and your brand does not hold a favourable recommendation role, or is not named at all. Three terms need pinning down first.

• A citation is one link an AI engine puts in its answer, pointing at a specific URL as a source for what it just said. In DiscoveredBy a citation is attributed to you when the link's registrable domain matches your project's domain, and nothing in that check reads the page.
• A mention is your brand named in the answer, with or without a link.
• A recommendation role is a label a model saves for each brand mention, such as top recommendation, alternative option, warning or caveat, or not recommended. The Sentiment documentation describes these as evidence to inspect, not proof the label is right.

The key terms page states the consequence plainly: you can be mentioned without being cited, and cited without being mentioned. The glossary entries for citation and brand mention give the short definitions.

Enlarge image

There is a fourth subtlety. Cited and named is not the same as recommended. An answer can name your brand as a caution ("popular, but limited integrations") while linking your page. That is why the role matters as much as the presence.

Why would an AI answer use your page and recommend someone else?

There is no way to read the engine's internal reasons, so the useful move is to list candidate explanations and test each against the saved answer. Treat every item below as a hypothesis, not a diagnosis.

1. The prompt was informational, and the page answers an informational claim. A question such as "how do compliance teams reduce contract review time?" invites an explanation. Your guide can support a sentence in that explanation without the answer ever turning to products. A prompt can carry a saved intent and buyer stage, and the Sentiment screen can filter by them, so you can check whether the prompts where this happens skew informational.
2. The cited page does not name your brand. If the page is a general explainer that never says who you are, there is nothing in the source for the answer to attach a brand to. The source detail screen checks the saved page text for your tracked brand names and aliases, which lets you see this directly.
3. The answer used your page for a fact, and chose products from elsewhere. The passage it credited might be a definition or a step list, while the product suggestions came from other sources in the same answer.
4. Your brand was named in a form the extraction did not resolve. A product name, abbreviation or misspelling that is not saved as an alias may not be linked to your brand family. Adding aliases is a documented control on the brands and sub-brands screen, though whether a name is recognised remains the extraction model's judgement.
5. The role is real, and it is unfavourable or lukewarm. The brand may be named, but as an alternative or a caveat. That is a framing problem, not an absence problem.
6. The citation is not what it looks like. For some engines the citation list is broader than the credited list. More on that below.

Notice that only some of these are about your content. The first is about the prompt, and the fourth and sixth are about how the record was built. Rule those out before rewriting anything.

How do you tell which pages the engine credited and which it only retrieved?

Check the engine first, because "cited" means different things across collection channels. The Retrieved vs cited screen shows, per source, how often an engine returned a page and how often the answer text attributed it.

The docs record several relevant behaviours:

• For Claude, every page its search returned is stored as a citation, so a Claude citation may be a page the engine only had in front of it. The Retrieved vs cited screen counts only a source the answer text credits.
• For Gemini (API) answers with no grounding links, every grounding source is stored as a citation.
• Perplexity Sonar records only sources its answer marks, and each carries a text position.
• Google AI Overviews, Google AI Mode and Gemini (app) report only the sources their answer cites, so there is nothing to compare citations against there. ChatGPT (app) also reports the pages it found, but its citations carry no positions for attribution, so the screen does not compare it yet.

Rates are computed for Claude, Perplexity and Grok. On the source detail page, an attributed answer shows up to three recorded passages, so you can read the sentence your page was credited for. For Grok that is the answer sentence carrying the citation; for Claude it is the passage Claude quoted from the page; for Perplexity it is the answer text just before the numbered marker.

That passage is the most useful single piece of evidence in this whole investigation. It tells you what your page was used for. For another angle on returned versus credited sources, see How often are returned sources also credited in the answer?.

Enlarge image

Step by step: the investigation

Work through one answer, then a small comparison set. The steps follow the order in which each depends on the one before.

Step 1: Fix the answer you are looking at

Record the exact prompt, engine and collection channel, date, and any country, language or persona filter. Keep collection channels separate, because DiscoveredBy does not average them: a channel is one provider, platform, surface and collection method, and API answers can differ from consumer-app responses.

Step 2: Read what the answer said about the claim your page supports

Open the saved answer and find the sentence next to your link. Where the source detail page shows a recorded passage, use it; for engines that record none, say so on the sheet. Ask what job the page did: a definition, a how-to step, a comparison criterion, or a product claim. A page credited for a general fact is doing a different job from a page credited for "the best tool for this".

Step 3: Check the prompt's intent

Look at the prompt's saved intent and buyer stage. An informational, awareness-stage question can end without any product being recommended, so an answer that cites you and recommends no one may not be an absence problem.

Step 4: Check whether your name is in the cited page

Open the URL's source detail page. It shows the saved copy of the page as our scraper last saved it, and it looks for your active brand and competitor names and aliases in that text. It also shows how many brands were checked, so an empty result is not read as a verdict when nothing was checked. It checks only the saved portion, and the copy may differ from the page the engine read when it answered.

Step 5: Record the framing of every brand named

On the Sentiment screen, open the Mentions tab of the Evidence card, find the records for this answer (each links to its prompt and collected outputs and shows an output ID), and read each brand's sentiment and recommendation role. If your brand is named, note whether the role is a top recommendation, an alternative option, a warning or caveat, or not recommended. If it is not named, there are no reasons to read for your brand: reasons come from the same extraction as mentions, and a brand that is only named, with no reason given, gets none.

Step 6: Read the stated reasons for the brands that were recommended

Brand reasons (available depending on your plan; see plans and limits) lists, per brand, the strengths and weaknesses the answer itself gave, each with a label from a fixed set of thirteen, a direction and a quote copied from the answer. A reason whose quote cannot be found in the saved answer is discarded. The label and direction are the extraction model's reading of the answer; the quote lets you check it. It does not show what caused the engine to prefer a brand.

Step 7: Compare with similar answers

One answer proves little. Pull answers to comparable prompts, on the same channel and across several days, and count how the four combinations fall. Fewer than 30 analysed answers is marked provisional in the product, so treat small counts as leads. The next section gives a way to tally them.

The citation-versus-recommendation investigation sheet

Copy this into a document or spreadsheet and fill in one copy per answer, then one summary per prompt group. The last column forces a claim-strength choice.

CITATION VS RECOMMENDATION SHEET

0. THE ANSWER
   Prompt (exact text):
   Prompt intent / buyer stage (as saved, or unclassified):
   Engine and collection channel:
   Date collected:
   Country / language / persona (if set):
   Link to the saved answer:
   Why this answer: (worst case / typical / random pick)

1. THE CITATION
   Cited URL:
   Domain (mine? subdomain of mine?):
   Rank in the answer:
   Did the engine only return it, or credit it in the text?  returned / credited / cannot tell
   Passage the answer credited it for (exact quote):
   Job the page did: definition / how-to step / criterion / product claim / other
   Other sources credited in the same answer (domain, page type):

2. THE PAGE
   Saved copy fetched? (date / not fetched / failed)
   Brand names checked and found in the saved text:
   Does the page name my brand anywhere?  yes / no / not checked
   Does it name a product or offer a next step?  yes / no

3. THE BRANDS NAMED
   Brand | Position | Sentiment | Recommendation role | Excerpt
   ---------------------------------------------------------------
   Mine:
   Competitor A:
   Competitor B:
   Was my brand named at all?  yes / no
   If named under an alias or variant, is the variant saved?  yes / no / n/a

4. STATED REASONS
   Brand | Label | Strength or weakness | Exact quote
   ---------------------------------------------------------------
   Mine:
   Competitor A:
   Competitor B:

5. HYPOTHESES (tick what the evidence supports, do not tick what it only fits)
   [ ] Informational prompt, informational use of my page
   [ ] Page never names my brand
   [ ] Page credited for a fact; products came from other sources
   [ ] Name variant not saved as an alias
   [ ] Brand named, but as alternative or caveat
   [ ] Citation reflects retrieval, not credit (engine-dependent)
   [ ] Other:

6. CLAIM STRENGTH FOR EACH FINDING
   Finding | Evidence | Claim strength: observed in this answer /
   repeated across N comparable answers / hypothesis only

7. NEXT ACTION (a test, not a fix)
   What I will change or check:
   How I will compare afterwards (same prompts, same channel, same window):

And a tally for the comparison set:

Use one row per prompt group and per channel, and check that the four counts add up to the number of answers.

Worked example

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

Quillstone sells document-review software. Its blog post "A checklist for reviewing vendor contracts" is cited in an answer to the prompt "How can a compliance team speed up contract review?" The answer recommends Brieflane and Clausewise and does not name Quillstone.

The team fills in the sheet.

• Citation: the URL is on Quillstone's domain, ranked second among four sources. The recorded passage is the sentence "Start by sorting clauses into standard and non-standard." The page's job: a how-to step.
• Page: the saved copy names "Quillstone" once, in the footer, and offers no product link in the body.
• Brands named: Brieflane (top recommendation), Clausewise (alternative option). Quillstone is not named.
• Prompt intent: saved as informational, awareness stage.
• Hypotheses supported: informational prompt; page credited for a how-to step; page barely names the brand.

The team then pulls ten answers to five similar informational, awareness-stage prompts on the same channel over the same 28 days. Ten is well below the 30 answers at which the product stops marking a sample provisional, so everything below is a lead.

The counts add up: 4 + 3 + 2 + 1 = 10. Quillstone is cited in 7 of the 10 (70%) and named in 5 of the 10 (50%). Of the seven answers that cite it, four (57%, 4 of 7) do not name it.

What the team can honestly write: "In 4 of 10 comparable answers on this channel, our checklist post is cited and our brand is not named. In the one answer we read closely, the credited passage was a how-to step and the page names us once." What they cannot write: "The engine ignores us because our post does not mention us." The evidence fits that idea; it does not establish it.

The next action is a test. They add a short, factual section to the checklist explaining how their product supports the same step, keep the prompts and channel fixed, and compare after a full window. If the naming rate moves, that is a lead worth repeating, not a proven cause. The habit of checking an edit against a fixed set is covered in did your content update help.

What this investigation cannot tell you

• Why the engine chose. A citation, a passage and a stated reason are observations of one answer. They do not reveal what the model weighed. Do not write "the engine ranks us lower because".
• What the page contained at answer time. The saved copy may differ from the page the engine read, and only the saved portion is searched for your name.
• That a name variant was missed. Brand recognition for new answers is the extraction model's judgement, even with aliases saved. A missing mention can be a recognition miss as easily as a real absence.
• Anything about consumer behaviour. These are collected answers, not what any one person saw. ChatGPT (app) and Gemini (app) answers are what those sites show someone who is not signed in.
• A complete list of citations. The Citations screen builds its tabs from the newest 500 citations in the window, so a busy window can hide older ones.
• Syndicated copies. A republished copy of your article on another publisher's domain is attributed to that publisher, not you.

A related and opposite case, where a brand is named and no source appears, is covered in what you can conclude when a brand is mentioned without a source.

Frequently asked questions

Is it good or bad when AI cites my site but does not name my brand?

It is a mixed signal. Being credited means an engine used your page as a source for something. Not being named means the answer did not attach that use to your brand, or named other brands for the product decision. Whether it matters depends on the prompt: for an informational question it may be unremarkable, and for a "best tool for X" question it deserves attention.

Does a citation mean the AI recommended my content?

No. A citation is a link the answer gives as a source for a statement, and the statement may be a definition or a background fact. A recommendation is a role the answer gives a brand. For some engines the citation list also includes pages the engine merely retrieved, so check the engine before reading any citation as credit.

Should I add my brand name to every informational page?

Not automatically. A page should name your brand where that is true and useful to the reader, and the aim is to test a hypothesis, not to force a mention. Make a change, keep the prompts and channel fixed, and compare over a full window before you repeat it elsewhere.

How many answers do I need before I trust the pattern?

The product marks a sample under 30 analysed answers as provisional, which is a sample-size cue and not a significance test. For a first pass, read a handful closely and treat the counts as leads. Build a fixed set of comparable prompts before you compare periods.

Why does my brand appear in the citation list but not in the mention list?

The two records come from different checks. A citation is matched by comparing the cited link's domain with your project domain. A mention comes from the model's reading of the answer text against your brand names and aliases. They can disagree, so read the saved answer to see which is right.

Next step

Open Citations to find answers that credit your domain, then use Sentiment and Brand reasons to read how the same answers framed each brand. The Retrieved vs cited screen shows the credited passage for the engines it compares. For the wider workflow, see why AI recommends your competitor, and for the feature overview, AI visibility. To run this investigation on your own answers, sign in or create an account.