Invisibility in AI answers has four distinct causes: no resolved entity, thin third-party evidence, wrong category placement, or retrieval failure. Each produces a different symptom and none share a fix.
You can tell them apart in an afternoon with four questions and no budget. The procedure is below, along with what to record and how to read the result. Buying a tool before you know which of the four you have is how teams spend a quarter fixing the wrong thing.
The complaint that starts most of these projects sounds the same every time: we asked ChatGPT for the best tools in our category and we were not in the list. That sentence describes a symptom, not a problem. Four very different failures produce it, and the fix for one is useless against the other three.
This page is the diagnostic. It is a living reference, revised as the engines change, and it deliberately stops short of the fixes so that you arrive at them knowing which one you need.
What does “invisible” actually mean here?
Before diagnosing anything, separate three states that get reported as the same thing. They look identical in a Slack message and they are not remotely the same problem.
Absent
Ask about you by name and the engine has little or nothing.
- What you see
- Vague, hedged, or a confident description of a different company.
- Points at
- Entity or retrieval.
Known but not listed
It describes you accurately, then omits you from every recommendation.
- What you see
- A good paragraph about you, and four competitors in the shortlist.
- Points at
- Evidence or category.
Listed but wrong
You are in the answer, described inaccurately.
- What you see
- Old pricing, a feature you removed, the wrong segment.
- Points at
- A stale source outranking your own.
Notice that only state A is invisibility in any literal sense. B and C are worse in practice, because the engine is actively saying something about you and it is not what you want said.
What are the four causes?
Underneath those states sit four causes. They are not a spectrum and a brand usually has exactly one dominant cause at a time.
1. No resolved entity
The engine cannot confidently say what your company is, or confuses you with a similarly named one. Everything downstream fails because there is nothing to attach it to.
State A
2. Thin third-party evidence
You are resolvable and described correctly, but almost everything written about you lives on your own domain, so you lose to competitors with denser independent coverage.
State B
3. Wrong category placement
You are known and well evidenced, but filed under a category buyers do not use for this question. You appear consistently for the wrong prompts.
State B
4. Retrieval failure
Your material exists and is good, and the retriever never reaches it. Blocked crawlers, pages that cannot be fetched, or content that does not survive being chunked.
State A or C
How do you tell which one you have?
What follows is the whole method. It takes an afternoon, needs no tooling, and produces a result you can defend to somebody who was not in the room. Run it in order, because each step narrows what the next one has to consider.
That is 90 answers, which sounds like a lot and takes about two hours. It is also the difference between a finding and an anecdote.
Ask three engines “What is [your company]?” and “Who is [your company] for?”. Record the full answer text, not a verdict. You are looking for whether the description is specific, whether it is correct, and whether it is about you rather than a similarly named company.
Vague or wrong here means cause 1. Stop and fix the entity before anything else.Open your own robots.txt and search it for the answer-engine crawlers, not just the training ones. Then ask an engine to summarise a specific page of yours by URL and see whether it can.
A blocked search crawler, or a page it cannot summarise, means cause 4. Also the cheapest thing on this list to fix.Ask ten unbranded questions a buyer would ask, phrased without your company name. Run each one three times. Record which brands appear and in what order.
Described well in step 1 but absent here means cause 2 or cause 3.Ask “What category of software is [your company]?” and then ask for the best tools in the category it named. If you appear there but not in the category you sell into, the placement is the problem.
Appearing under the wrong label is cause 3. Absent everywhere but described well is cause 2.# Step 1, identity. Run on three engines.
What is [company]?
Who is [company] for?
# Step 3, buying questions. Unbranded. Your name appears nowhere.
Best [category] tool for a [size] [industry] team?
What should I look for when choosing a [category] tool?
Which [category] tools handle [specific job] well?
# Step 4, placement.
What category of software is [company]?
What are the best tools in [whatever it just answered]?What to record, and in what shape
The value of the exercise is entirely in the discipline of the record. One sheet, one row per run.
date,engine,prompt_id,run_no,brands_named,name_order,cited_urls,answer_text
2026-08-13,chatgpt,Q1,1,"You;Vendor A;Vendor B","3rd","a.com;b.com",""
2026-08-13,chatgpt,Q1,2,"Vendor A;Vendor C","n/a","a.com;c.com",""
2026-08-13,chatgpt,Q1,3,"You;Vendor A","2nd","a.com","" Specific, reproducible, and it points at one cause. Somebody can re-run it next month and compare.
Which prompt, which engine, how many runs, and known in what sense? This is the sentence that starts a quarter of misdirected work.
What does each cause look like up close?
Cause 1: no resolved entity
The engine does not have a confident idea of what you are. It hedges, describes you generically, or attaches facts belonging to another company with a similar name.
The tell is asymmetry: it can talk about your category fluently and about you barely. Fixes live in the entity layer, and Google is unusually explicit that this is signal-driven rather than declared. Its site names documentation states the process is automated and that WebSite structured data is “most important”, ahead of og:site_name, title elements and headings. The sameAs property on schema.org’s Organization type does the same job for identity, defined as a URL that “unambiguously indicates the item’s identity”.
Confirmation, though, comes from outside. Wikidata’s notability policy admits an item that can be “described using serious and publicly available references”, which is a fair statement of what an engine is looking for too.
Cause 2: thin third-party evidence
You are resolvable, described accurately, and still never shortlisted. Almost everything an engine can find about you lives on rivarise.com, or whatever your equivalent is, and self-description is the weakest evidence there is.
This is the most common cause among well-run B2B companies, and their competence is what causes it. A team that writes excellent documentation, keeps a detailed pricing page and publishes twice a week has answered every question about itself on its own domain, which feels like thoroughness and reads to a retriever as a single source repeating itself.
The confirming test is a counting exercise rather than a judgement. Take the ten buying questions from step 3, collect every URL cited across all runs, and sort them by domain.
Every cited URL is yours. The engine has one source for you, and that source is you describing yourself.
Review platforms, comparison articles, community threads, a directory or two. Several independent sources agreeing, which is what an engine treats as confidence.
A team that publishes twice a week on its own domain has answered every question about itself in one voice. That reads as thoroughness internally and as a single source repeating itself to a retriever.
It is also the slowest cause to fix, which is exactly why nobody should start a two-quarter evidence programme on a hunch.
Cause 3: wrong category placement
You are visible, just filed somewhere your buyers are not looking. A product sold as “revenue intelligence” that engines consistently file under “CRM add-ons” will appear for the wrong questions and be absent from the right ones.
This one is easy to miss because internally it looks like success. The engine describes you accurately, uses your own words back at you, and names you in answers. It is only when you compare the questions you appear for against the questions your buyers actually ask that the mismatch shows up, which is why step 4 asks the engine to name your category rather than asking it about the category you assume you are in.
The fix is positioning work done in public. Pick the category buyers search, then use it identically on your own site, in every directory profile, in press, and in the first sentence of your own description everywhere it appears. Consistency is the mechanism, not repetition: an engine resolves a category the same way it resolves an entity, by finding independent sources that agree. That makes it slower than a website rewrite and considerably faster than an evidence programme.
Cause 4: retrieval failure
Everything is right and the retriever never gets there. Two things cause it more than anything else.
OpenAI and Perplexity both separate their answer crawler from their training crawler. A blanket rule blocks both, and the first one is what decides whether you can be named.
Engines retrieve passages, not documents. A section that only makes sense after reading the three above it rarely gets retrieved, and never gets quoted cleanly.
One correction worth making while you are here: robots.txt is not the tool for hiding a page. Google states it “is not a mechanism for keeping a web page out of Google” and that a disallowed page “can still be indexed if linked to from other sites”, per the robots.txt introduction. It governs crawling. Use noindex if you actually want a page gone.
What should you not conclude from a single check?
The single most expensive error in this whole exercise is treating one run as a result. Answers vary between identical runs, so a check that returns nothing tells you nothing on its own.
You were absent from run two and present in runs one and three. Nothing changed except the sampling.
You appeared once, in the run somebody screenshotted for the board. Your actual rate across thirty runs is under a fifth.
A change made on Tuesday looks like it worked on Friday. Without a volatility band you cannot separate the fix from the weather.
There is also a timing trap. If step 2 finds a blocked crawler and you fix it, do not re-run the diagnostic the same evening. OpenAI documents that a robots.txt change can take around 24 hours to register on their side, so a same-day re-test measures your patience rather than your fix.
What do you do once you know?
Each cause routes somewhere different, and the ordering between them is not negotiable. Entity work comes first regardless of which cause you have, because a resolved entity is what the other three attach to.
Organization markup with sameAs pointing only at references that exist, and one canonical description you use in every profile.The diagnostic is worth an afternoon because the cheapest of these fixes is done by Friday and the most expensive is a quarter of somebody’s roadmap.
What if the diagnostic is ambiguous?
Two causes can be live at once, and the record usually shows it. The rule is to act on the cheapest one first and re-measure before touching the expensive one, because a fix in the access layer can change what the entity layer looks like.
If step 1 gives a decent description but step 2 finds a blocked crawler, fix the crawler, wait a day, and run steps 1 and 3 again. Roughly a third of what looks like an entity problem on a blocked site turns out to have been retrieval all along, because the engine was working from whatever third-party material it could reach rather than from you. That sequencing costs you two days and can save you a quarter.
If nothing is conclusive after a full pass, the honest answer is that your prompt set is wrong rather than your brand being unusual. Ten questions phrased the way your team talks about the category will produce ten uninformative answers. Ask somebody who bought recently how they searched, and use their words instead of yours.
None of this requires believing anything about AI search that you cannot check. Google’s own position is that no AI-specific optimization exists for its surfaces, and its helpful content guidance still asks the same questions it asked before any of this: is it self-evident who authored the content, and does it demonstrate first-hand expertise.
Run the four questions. Write down what actually came back. Whichever of the four causes the record points at, you will have spent an afternoon instead of a quarter finding out.