Quick answer
AI brand detection is not just a checkbox that says "the brand was mentioned."
The useful question is whether the system matched the right entity in the right answer, with the right competitor set, for the right prompt group. A brand-detection setup should catch real mentions, avoid false positives, recognize common aliases, separate subsidiaries or products, and preserve the raw answer that created the metric.
Use this QA loop before trusting visibility totals:
| QA step | What to check | Why it matters |
|---|---|---|
| Define the entity | Official name, product names, old names, acronyms, misspellings, and domains. | A narrow setup misses real mentions. A broad setup counts noise. |
| Define competitors | Direct competitors, category leaders, regional alternatives, and common abbreviations. | Share of Voice is only meaningful against a clean comparison set. |
| Read raw answers | Inspect the exact text, not only the mention count. | A matched word may be a different entity or a generic phrase. |
| Review citations | Check whether the answer cites your domain, another source, or no visible source. | A mention without source evidence has a different fix than a citation gap. |
| Label errors | False positive, false negative, wrong competitor, stale fact, or prompt mismatch. | Each error needs a different owner and fix. |
| Re-measure | Re-run the same prompt group after setup or source fixes. | The dashboard should improve because the evidence improved, not because the setup hid a problem. |
For ReachLLM teams, the operating model is: configure aliases, run prompts, inspect raw answers, classify detection errors, fix the brand source of truth, and re-measure the same prompt group.
What the Otterly source gets right
The scheduled source for this run is Otterly's help article about brand detection. It is useful as a market signal because it makes two practical points: brand reports depend on detecting brand names inside AI responses, and teams should define brand-name variations and competitor-name variations in report settings.
The page also says Otterly updated its brand detection on November 27, 2025, and describes the newer approach as primarily text based. Its example is simple: if a brand defines "Coca Cola" and "Coca-Cola," the system looks for exact matches of those terms in AI responses.
That is a helpful reminder. Brand detection depends heavily on setup quality.
The ReachLLM addition is the QA layer. A visibility team should not stop at entering a brand name. It should test whether the setup works across aliases, product lines, subsidiaries, competitor names, misspellings, and prompts where a word can mean something else.
Why brand detection errors distort the whole report
Most AI visibility metrics depend on brand detection.
| Metric | How detection errors distort it |
|---|---|
| Visibility Score | Missed aliases undercount visibility; false positives inflate it. |
| Share of Voice | Dirty competitor names make the comparison unfair. |
| Average Rank | Wrong entity matches can move a brand into a list where it did not appear. |
| Citation rate | A cited page may mention the brand, but the answer may not recommend it. |
| Sentiment | A negative phrase about another entity can be assigned to the wrong brand. |
| Trend lines | Setup changes can look like visibility movement if they are not annotated. |
That last point is easy to miss. If you add five aliases today, tomorrow's Visibility Score may rise because the detector improved, not because buyers are seeing a stronger answer. Treat setup changes as measurement changes and annotate them in the report.
Step 1: build the brand dictionary
Start with a tight dictionary. Add variants because real answers are messy, but do not add every word a human might associate with the company.
| Entry type | Include | Be careful with |
|---|---|---|
| Legal company name | Registered company and public trading name. | Legal suffixes that never appear in buyer answers. |
| Product name | Product and platform names buyers use. | Generic product words that match unrelated answers. |
| Acronym | Acronyms the market actually uses. | Acronyms shared with other companies or concepts. |
| Misspelling | Common misspellings found in raw answers or search data. | Rare typos that create noisy matches. |
| Domain | Primary domain and important subdomains. | Domains mentioned as citations but not as brand recommendations. |
| Old name | Former names that still appear in sources. | Retired names that should be corrected rather than celebrated. |
| Subsidiary or parent | Only when buyers treat it as the same buying entity. | Separate brands that need separate tracking. |
For a brand like ReachLLM, the dictionary should cover the public brand and domain. It should not blindly count every mention of "LLM," "GEO," or "AI visibility" as a brand mention. Those are category terms.
Step 2: set competitor names with the same discipline
Competitor detection needs the same QA as brand detection. If the brand setup is precise but competitors are loose, Share of Voice becomes unreliable.
For each competitor, capture:
| Competitor field | Example use |
|---|---|
| Official name | The name used on their site and pricing pages. |
| Common short name | The name buyers use in prompts and forum discussions. |
| Product name | Useful when the company and product names differ. |
| Domain | Helps review citations and source trails. |
| Exclusion note | Terms that should not count because they mean something else. |
Then test the setup on raw answers. A competitor named after a common word, acronym, or category phrase needs stricter matching than a unique brand name.
Step 3: separate four detection cases
Do not group every detection issue under "the brand was missed." Use four cases:
| Case | What happened | First action |
|---|---|---|
| True positive | The answer clearly names the right brand or product. | Keep the match and inspect rank, source, sentiment, and citation context. |
| False positive | The detector counted a different entity, generic phrase, or unrelated citation. | Tighten aliases, add exclusions, or require phrase/context matching. |
| False negative | The answer names the brand through a valid alias but the detector missed it. | Add the alias and annotate the setup change. |
| Ambiguous match | The word could refer to the brand or another entity. | Read the full answer and cited sources before counting it. |
This prevents teams from arguing about dashboard totals without looking at the evidence.
Step 4: audit raw answers by prompt group
Brand detection should be reviewed by prompt group, not only in aggregate.
| Prompt group | Detection risk |
|---|---|
| Branded fact prompts | Misspellings, old names, wrong product names, stale facts. |
| Category discovery prompts | Competitors appear without the brand, or aliases collide with category terms. |
| Comparison prompts | Product names and company names get mixed. |
| Pricing prompts | Plan names can be mistaken for brand mentions. |
| Local or regional prompts | Local subsidiaries, branches, or unrelated same-name businesses can appear. |
| Support or sentiment prompts | Negative text may refer to a competitor or a different product line. |
Review at least a sample of raw answers after any setup change. If one prompt group has many false positives, the total visibility number is not ready for reporting.
Step 5: inspect citations before assigning the fix
Brand detection tells you whether the answer named the brand. Citations tell you what source evidence may have shaped the answer.
| Detection result | Citation pattern | Likely diagnosis |
|---|---|---|
| Brand mentioned and owned domain cited | Owned source is reachable and relevant. Review answer quality and sentiment. | |
| Brand mentioned but third-party source cited | The answer may rely on external proof. Check whether the source is accurate. | |
| Brand mentioned with no visible citation | Treat as a clue, not a source-backed win. Preserve the answer and re-check. | |
| Competitor mentioned and cited | Inspect the source that validated the competitor. | |
| Brand missed but owned page cited | The page may answer the topic without making the brand role clear. | |
| Brand counted from a citation only | Do not treat a URL citation as a recommendation unless the answer text names the brand. |
Google's AI feature guidance says the same Search fundamentals remain relevant for AI Overviews and AI Mode, and that there are no additional technical requirements or special AI-only schema needed for inclusion (Google Search Central). It also says eligible pages need to be indexed and eligible for a snippet. That supports technical hygiene, but it does not mean a detected brand mention is automatically a recommendation.
Step 6: route errors to the right owner
A brand-detection problem can be a configuration issue, source issue, content issue, or product-positioning issue.
| Error | Likely owner | Better fix |
|---|---|---|
| Alias missed | AI visibility or marketing operations | Add the alias, document the setup change, and re-run the sample. |
| Generic term counted | AI visibility or data owner | Remove the alias or require stricter matching. |
| Competitor alias missing | Product marketing | Add the competitor variation and rebuild Share of Voice comparison. |
| Old brand name appears | Product marketing or comms | Correct source-of-truth pages and reachable profiles. |
| Brand omitted from category prompts | Product marketing, content, or PR | Improve category proof, owned pages, and cited third-party sources. |
| Wrong entity cited | SEO, web, or operations | Fix structured source data, profiles, internal links, and external profiles where possible. |
| Setup change moved the score | Reporting owner | Annotate the report so leadership does not mistake setup QA for market movement. |
Do not hide a real visibility gap by deleting a competitor, prompt, or alias. Clean measurement should make the work clearer, not easier to spin.
Step 7: keep technical eligibility separate from detection
Detection QA asks whether the system counted the right text. Technical eligibility asks whether important pages can be discovered and understood.
Check:
| Technical check | Question |
|---|---|
| Robots access | Are important pages blocked from search or AI retrieval crawlers that matter to the workflow? |
| Indexability | Are priority pages free of accidental noindex or restrictive snippet controls? |
| Canonical tags | Do canonical URLs point to the intended source page? |
| Visible text | Are brand facts available as text, not only images or client-only widgets? |
| Sitemap and internal links | Can crawlers and users find the page? |
| Structured data | Does schema match visible content and real business facts? |
| Source-of-truth files | Are llms.txt, homepage, platform page, pricing, docs, and profiles consistent? |
OpenAI's crawler documentation says OAI-SearchBot is used to surface websites in ChatGPT search features, and that sites can manage search opt-outs through robots.txt (OpenAI Developers). Allowing a crawler supports eligibility. It does not guarantee a mention, citation, rank, or sentiment.
Where ReachLLM fits
ReachLLM is useful when brand detection needs to become trusted measurement and shipped work.
The platform tracks prompts across enabled AI systems, preserves raw answers, compares competitors, and measures Visibility Score, Share of Voice, Average Rank, citation rate, sentiment, source data, and history. It also connects findings to GEO audits, source intelligence, brand-fact review, content updates, website fixes, structured data, llms.txt, PR outreach, integrations, agent workflows, governed approvals, and managed execution.
Use ReachLLM when:
| Need | How ReachLLM helps |
|---|---|
| The team needs defendable mention data. | Raw answers, prompt groups, competitors, citations, and history stay connected. |
| Brand aliases or competitor names are messy. | Prompt-level review separates setup errors from real visibility gaps. |
| AI answers use old or wrong brand facts. | Brand-fact review turns stale claims into source-of-truth corrections. |
| Competitors are detected more often. | Source intelligence shows which pages, domains, and proof paths are shaping the answer. |
| The fix requires execution. | GEO audits, content, website changes, schema, llms.txt, PR, and managed execution can be routed from the same evidence. |
If a team only needs a one-time manual spot check, a spreadsheet can work. If detection errors affect reporting, client decisions, or roadmap priorities, the team needs raw-answer evidence, source review, and re-measurement in the same workflow.
A brand-detection QA note template
Keep one note per prompt group:
| Field | Example |
|---|---|
| Prompt group | Category discovery prompts for US mid-market buyers. |
| Brand dictionary checked | Official brand, product name, domain, common misspelling. |
| Competitors checked | Three direct competitors and two regional alternatives. |
| Detection issue | False negative: answer used product name but not company name. |
| Source evidence | Third-party roundup cited the product name only. |
| Fix | Add product alias and update owned category page to connect product and company name. |
| Owner | Product marketing owns page copy; AI visibility owner updates tracking setup. |
| Re-check | Same prompt group after setup update and next scheduled run. |
This note is more useful than a screenshot of a score with no explanation.
What not to do
- Do not treat a citation URL as a brand mention unless the answer text names the brand.
- Do not count broad category terms as brand aliases.
- Do not delete competitors to improve Share of Voice.
- Do not merge product, parent, subsidiary, and partner mentions without a written rule.
- Do not report a visibility jump without noting alias or competitor setup changes.
- Do not claim schema,
llms.txt, crawler access, or one page update guarantees detection. - Do not copy competitor help docs or source pages that currently shape the market.
FAQ
What is AI brand detection?
AI brand detection is the process of identifying when an AI-generated answer names a specific brand, product, competitor, or entity. Useful detection preserves the prompt, raw answer, platform, matched term, citation context, competitor set, and date.
Why do aliases matter in AI visibility tracking?
Aliases matter because AI answers may use a product name, old name, acronym, misspelling, domain, or shortened brand name. Missing aliases can undercount visibility, while overly broad aliases can create false positives.
How should teams QA AI brand detection?
Review raw answers by prompt group, compare matched terms against the brand and competitor dictionaries, label false positives and false negatives, inspect citation context, annotate setup changes, and re-run the same prompts before reporting movement.
Is a brand mention the same as a citation?
No. A brand mention means the answer names the brand. A citation means the answer exposes a source URL. A brand can be mentioned without being cited, and a domain can be cited without the answer recommending the brand.
How does ReachLLM help with brand-detection QA?
ReachLLM connects prompt tracking, raw answers, competitors, citations, source intelligence, sentiment, brand-fact review, Visibility Score, Share of Voice, Average Rank, GEO audits, content updates, website fixes, llms.txt, PR outreach, and re-measurement in one workflow.
Sources reviewed
- Otterly Help Center, "Brand Detection: How is OtterlyAI detecting Brands in AI Responses?": https://help.otterly.ai/brand-detection
- Google Search Central, "AI features and your website": https://developers.google.com/search/docs/appearance/ai-features
- Google Search Central, "Creating helpful, reliable, people-first content": https://developers.google.com/search/docs/fundamentals/creating-helpful-content
- OpenAI Developers, "Overview of OpenAI Crawlers": https://developers.openai.com/api/docs/bots