Why AI recommends your competitor: how to investigate the answer

When an AI answer recommends your competitor instead of you, the way to find out why is to stop looking at scores and read one saved answer, keeping four questions apart: did your brand appear, how was it framed, what reasons did the answer state, and which sources did it credit. Each question has different evidence and a different next step. What you learn is a description of that answer. It does not reveal why the model produced it, so treat every finding as a hypothesis to test, not a cause to fix.

In short

• Investigate one saved answer before you look at trends. A trend tells you where to look; an answer tells you what happened.
• Keep four questions separate: appearance, framing, stated reasons, credited sources. Mixing them is how teams end up fixing the wrong thing.
• A stated reason is what the answer said, not what drove the model. A cited source is a page the answer linked, not proof it shaped the recommendation.
• Move from one answer to a small set of comparable answers before you decide anything.
• Write each finding at the strength the evidence supports. The worksheet below has a claim-strength column for that.

Why start with one answer, not a dashboard?

A saved answer is the only place where all four kinds of evidence sit side by side. Aggregate numbers such as a visibility score are useful for noticing a problem, but they cannot say what an answer said about you or which pages it linked.

Reading one answer closely takes little time. It tells you whether your problem is absence (you are not named), position (named, but as an also-ran), reasons (the answer favours a rival on something you also do well), or sources (the pages linked are ones you have no presence in). Those problems need different responses, which is why they stay separate.

Enlarge image

Which four questions should you separate?

Ask them in this order, because each one depends on the one before it. Definitions first, since the words get used loosely.

• A mention is your brand named in an AI answer. A citation is a link in the answer pointing at a specific URL as a source. You can be mentioned without being cited, and cited without being mentioned. DiscoveredBy's key terms page draws the same line.
• Framing is how the answer treats a brand: its sentiment (positive, neutral or negative) and its recommendation role, such as top recommendation, alternative option, warning or caveat, or not recommended. In DiscoveredBy these are labels a model saves for each brand mention, and the Sentiment page describes them as evidence to inspect, not proof that the judgement is right.
• A stated reason is a reason the answer itself gives for its view of a brand, such as "cheapest option" or "limited integrations", tagged with a label and a direction (strength or weakness).
• A credited source is a page the answer linked as support, which is a citation. (For some engines, such as Claude, every page the search returned is stored as a citation, so check the engine before reading a citation as "credited".)

1. Did your brand appear?

Check whether your brand is named at all, and where. If it is absent, there is no framing or reasons to read for your brand, only for the brands that were named. If you are named, note the order of brands and whether you are in the main recommendation or an aside.

2. How were you framed?

Record the sentiment and the role for your brand and for each competitor named. Roles matter more than tone: an answer can be warm about you ("a solid tool") and still place you as an alternative behind a top recommendation.

3. What reasons did the answer state?

Reasons are where an answer explains itself. Look for the specific things it says each brand is good or weak at, and note whether each is presented as a reason to choose the brand or a reason for caution. Copy the exact phrase, because the phrase is the evidence.

This is where teams overreach, so hold one line firmly: stated reasons are the answer's own explanation in its own words. They do not reveal what the model weighed internally or why one brand came first. They show which arguments the answer chose to voice.

4. Which sources did it credit?

List every link the answer gave: the domain, the type of page (your site, a competitor's, a review site, a publication, a forum), and the claim it seems attached to. Then note whether any is yours. One of your pages cited in an answer that still recommends a rival is a different finding from none of your pages being cited.

What does DiscoveredBy record for each question?

If you use a monitoring tool, check what it stores for an answer before relying on it. This is what the DiscoveredBy docs describe.

For stated reasons, the Brand reasons screen shows, for your brand and each active tracked competitor, how many answers gave each reason as a strength or a weakness, with the quote behind every count. Each reason has a label (one of thirteen, such as pricing, features, ease of use, integrations, support, trust or security, with "Other reason" as the catch-all), a direction, and a quote copied from the answer. A reason whose quote cannot be found in the saved answer text is discarded, and a brand has at most three reasons in one answer. The label and direction are the extraction model's reading of the answer; the quote is there so you can check that reading.

One limit matters here: answers analysed before the feature shipped are not re-analysed, so they are counted as "not analysed" and never as answers with no reasons.

For framing, the Sentiment screen lists stored mention excerpts with sentiment and recommendation role, and a link to the prompt and its collected outputs. For sources, the Citations screen lists one row per citation with its rank in that answer and a link to the source. A citation is attributed to you when its registrable domain matches your project's domain (a subdomain of yours counts as you). Nothing in that check reads the page, so a syndicated copy of your article on another publisher's domain is attributed to that publisher.

One caution on the Citations screen: its tabs are built from the newest 500 citations in the window, so a busy window can hide older ones.

Enlarge image

The diagnosis worksheet

Copy this into a document or spreadsheet and fill it in for one saved answer. Each row asks for a finding, the evidence, and the strongest claim that evidence supports.

DIAGNOSIS WORKSHEET: one saved answer

0. WHAT I AM LOOKING AT
   Prompt (exact text):
   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. APPEARANCE
   Was my brand named?                     yes / no
   Brands named, in order:
   My position in the order:
   Named in the main recommendation or an aside?

2. FRAMING (one line per brand named)
   Brand | Sentiment | Recommendation role | Excerpt
   -----------------------------------------------------
   Mine:
   Competitor A:
   Competitor B:

3. STATED REASONS (one line per reason)
   Brand | Label | Strength or weakness | Exact quote
   -----------------------------------------------------
   Mine:
   Competitor A:
   Competitor B:
   Reasons a rival gets that I could also claim truthfully:
   Reasons against me that are wrong, outdated, or true:

4. CREDITED SOURCES (one line per link)
   URL | Domain | Page type | Claim it sits next to | Mine? | Rank in answer
   -----------------------------------------------------
   Sources that mention my brand:
   Sources in my category that omit my brand:
   Any of my own pages cited?              yes / no

5. WHAT I CAN AND CANNOT SAY
   Observed in this answer (fact):
   Pattern to check across more answers (hypothesis):
   Not supportable from this evidence:

6. NEXT CHECK
   Comparable answers to open (same prompt, other days or engines):
   One hypothesis to test:
   What would change my mind:
   Owner and date:

How to fill in section 5

Section 5 is the one people skip and the one that keeps the worksheet honest. Use this table to set the strength of each claim.

From one answer to a pattern

One answer can be a fluke. Before you change anything, open a handful of comparable answers and check whether the same findings recur.

Keep the population comparable: the same prompt, the same engine and collection channel (one provider, platform, surface and collection method), and a window of days you choose in advance. DiscoveredBy does not pool channels, and the same discipline applies by hand.

Then look at the reason counts. On the Brand reasons screen, a rate is answers with that reason and direction divided by answers analysed. A family with fewer than 30 analysed answers is marked provisional, which is a sample-size flag rather than a confidence interval. Treat a provisional rate as a lead, not a result.

If sources stand out, note that an engine may return more pages than its answer credits (see Retrieved vs cited); a page returned but not credited is a different problem from one that never appeared. The follow-up post your page was retrieved but not cited goes into that case.

Worked example: Quillstone and the compliance-team prompt

Quillstone sells document-review software to mid-sized legal and compliance teams. Its content lead notices that an engine keeps recommending Brieflane for the prompt "What is the best document review software for a mid-sized compliance team?" She opens one saved answer and fills in the worksheet.

1. Appearance. The answer names three brands in this order: Brieflane, Clausewise, Quillstone. Quillstone is named, so this is not an absence problem. It is third of three.

2. Framing. Brieflane is a top recommendation with positive sentiment. Clausewise is an alternative option, neutral. Quillstone's role is warning or caveat, with neutral sentiment. The role is the finding: Quillstone is present but flagged with a caveat rather than recommended.

3. Stated reasons. The saved answer gives these reasons, each with a quote she copies:

Two things stand out. The integrations reason appears on both sides: a strength for Brieflane, a weakness for Quillstone. And Quillstone has a strength the answer stated (security) that the answer did not present as enough to recommend it.

4. Credited sources. The answer links four pages: a Brieflane integrations page, a legal-technology publication's roundup, a review site listing, and a Clausewise pricing page. None is on Quillstone's domain. The roundup names Brieflane and Clausewise but not Quillstone.

5. What she can say. In this answer, Quillstone was named third, framed with a caveat about integrations, and none of the four linked pages were its own. She cannot say why Brieflane came first, or that integrations decided it.

6. Next check. She opens the other answers collected for this prompt on the same channel over the last 28 days. Ten were analysed for reasons. In six of the ten (60%) integrations is given as a strength for Brieflane. In four of the ten (40%) integrations is given as a weakness for Quillstone. Security is a strength for Quillstone in three of ten (30%). With only ten analysed answers, every rate is provisional (fewer than 30 analysed answers), so she records it as a pattern worth testing, not a finding.

Her hypothesis: the answers repeat an integrations weakness that Quillstone's public pages do not address. If its connectors are documented only behind a login, a public integrations page could give answers something to cite. That is a change to try and measure, with no promise attached. The post turn an AI citation gap into a content brief your writer can use shows how to hand that hypothesis to a writer.

Enlarge image

What this investigation cannot tell you

An answer describes itself; it does not reveal the model. These are the limits to state plainly in any write-up.

• Stated reasons are not internal causes. A reason quote is the answer's own explanation in its own words. The Brand reasons page says the quote lets you check the extraction model's reading and does not show what caused an engine to rank a brand.
• Citations are not proof of influence. A cited page was linked in the answer. That does not establish that the page drove the recommendation, and a brand can be recommended with no link at all.
• One answer is one draw. The same prompt can produce different answers on different days and channels.
• Extraction is a model's reading. Sentiment, roles, reason labels and directions are saved model classifications. Sample the quotes to check them.
• Missing is not neutral. An answer with no reasons for your brand may simply not have been analysed.
• Weaknesses are only the ones answers volunteer. Tracked prompts do not usually ask for downsides, so the weaknesses you see are the ones the answers happened to give.

The mistakes that follow: fixing a page because a competitor's page was cited, treating the most common reason as the cause, and drawing conclusions from a provisional sample.

Frequently asked questions

Why does AI recommend my competitor and not me?

You cannot know the internal cause. You can observe whether you were named, how you were framed, what reasons the answer stated and which sources it linked. Those point to hypotheses, such as a recurring weakness or a missing source, that you can test.

Should I ask the AI engine directly why it recommended someone else?

That produces another generated answer, not a window into the earlier one. Treat it as one more observation. The reasons stated in the original saved answer are better evidence.

How many answers do I need before I act?

There is no universal number. Review a set of comparable answers (same prompt, engine and channel, over a window set beforehand) and see whether the same finding recurs. DiscoveredBy flags rates based on fewer than 30 analysed answers as provisional, so treat those as leads.

What if my brand is named but the answer does not recommend it?

That is a framing and reasons problem, not an absence problem. Read the role your brand received and the reasons stated for it. If a stated weakness is factually wrong, that is a fact-correction task; if it is true, it is a product or positioning question.

Does being cited mean I was recommended?

No. A citation means the answer linked one of your pages as a source. You can be cited and still framed as an alternative or a caveat, and you can be recommended without any of your pages linked. Check both the framing and the sources.

Can I do this without a monitoring tool?

Yes, for a few answers. Save the answer text, the links and the date, and fill in the worksheet by hand. A tool helps when you need many comparable answers and saved history.

Next step

If you already track prompts in DiscoveredBy, open the saved answer behind a competitor recommendation, follow the link from a Brand reasons quote to the prompt and its collected outputs, use the Citations screen for the links that answer gave, and fill in the worksheet. The brand perception and sources pages describe the two sides of this workflow. Brand reasons availability depends on your plan; see plans and limits.

Start tracking your first prompts