Because at least one of three things is missing. An AI engine tends to name a company when it can reach that company’s site, resolve the name to one clear entity, and find outside sources that agree. Miss one and you get one of two outcomes, both bad. Either the engine does not mention you at all, or it fills the gap with whatever it can reach, which is how we ended up being described as an operations firm with a competitor recommended in our place.
Orion & Taurus sells this work, so read this the way you would read any explainer published by a vendor. Every outside figure below is linked to its primary source, and every source was opened and read by hand on August 31, 2026. The evidence about our own company is our own, it was gathered in single runs rather than repeated ones for reasons explained below, and you can run the same test on yourself in half an hour.
One caveat up front, because you will find the exception yourself if we do not. The first condition is not strictly required. The New York Times blocks GPTBot, OAI-SearchBot, ClaudeBot, Claude-SearchBot and PerplexityBot in its robots.txt, checked August 31, 2026, and ChatGPT and Claude will still tell you exactly what The New York Times is. But notice what is carrying it: the Times is the most resolvable and most corroborated news entity in the English language, so conditions two and three are not being skipped, they are maxed out. Enough of those two can buy you out of the first. Almost nobody reading this has that. This is a working model built from audits, not a documented ranking system.
What does AI invisibility actually look like?
We ran our own firm through the engines on August 31, 2026, in fresh sessions with no account history, so the results are what a stranger sees rather than what our signed-in account sees.
We asked ChatGPT what Orion & Taurus is and told it to search the web and cite its sources. It returned roughly 2,700 characters about constellations. Orion the Hunter, Betelgeuse, Rigel, the Orion Nebula, Taurus the Bull. It cited NASA. A string search of the full response found zero occurrences of our domain, of “AI visibility”, or of the founder’s name. Be precise about what happened there: the engine did not get our company wrong, it did not know a company was being asked about.
Then we asked Google’s AI Mode the same question, with the prompt loaded in our favour by naming the category outright: what is Orion & Taurus, the AI visibility consultancy? It answered that we are “actually a Fractional COO (Chief Operating Officer) firm, rather than a dedicated AI visibility consultancy”, sourced that from a second domain we own, and then suggested the reader might be thinking of a competitor, which it named.
Across every engine that answered, our main website was cited zero times. The sources the engines drew on included that second domain of ours, a competitor, two unrelated third-party articles about AI visibility, and NASA. Not one independent source describing us.
Two honest limits. Of the five AI engines we track, ChatGPT, Perplexity, Google AI Mode, Copilot and Claude, only two answered in a clean logged-out session, and one of those hit a login wall after a single question. Because of those walls we ran each prompt once rather than twice, which makes this a snapshot rather than a rate, by the same standard we set for you further down. And we removed those signals from the second domain on August 31, 2026, the day we ran the test. Whether the engines re-resolve, and how long it takes, is what we will report next.
Why did this become urgent in 2026?
Two things changed. Fewer people click through, and a growing share of the interest in your site is automated and AI-purposed. If the vocabulary is new to you, we cover it in what is answer engine optimization, which is the same work under the label most of the industry uses.
On the click side, the Pew Research Center tracked 68,879 unique Google searches from 900 U.S. adults on Ipsos’ KnowledgePanel Digital, using browsing data from March 2025. On visits where Google showed an AI summary, people clicked a traditional result 8% of the time. Without a summary, 15%. A link inside the summary was clicked on just 1% of visits to pages carrying one (Pew Research Center, July 22, 2025, checked August 31, 2026). That is association rather than cause, since searches that trigger a summary may differ in intent, and it is still the most rigorous dataset of its kind we could find.
On the machine side, Cloudflare reports that 52% of the crawler requests it classifies by purpose were for AI training as of June 2026, up from 22% in spring 2025, with a further 36% from mixed-use crawlers blending search, agents and training (Cloudflare, July 1, 2026, checked August 31, 2026). That measures what a subset of crawlers wanted on Cloudflare’s network. It is not a count of your site’s traffic, and it is not proof that machines outnumber human readers. The point that survives is narrow and still worth acting on: AI-purposed crawling is now a large and growing share of automated attention, and none of it appears in your analytics.
Can the engine reach your site?
This is the condition most owners assume is fine, and it is the one most often quietly broken.
AI companies do not run one crawler. OpenAI documents four: GPTBot for training data, OAI-SearchBot to surface sites in ChatGPT’s search features, ChatGPT-User for actions a person triggers inside ChatGPT, and OAI-AdsBot, which only visits pages submitted as ads. Only two are robots.txt control tokens, GPTBot and OAI-SearchBot, and OpenAI states the two settings are independent. ChatGPT-User is different: because a person initiates it, OpenAI says robots.txt rules may not apply, and directs site owners to OAI-SearchBot for search opt-outs (OpenAI, checked August 31, 2026).
Anthropic documents three and treats them more uniformly: ClaudeBot for training, Claude-User for user questions, Claude-SearchBot for search quality, all of which it says honour robots.txt (Anthropic, checked August 31, 2026).
The distinction that costs money is training versus search. Blocking a training crawler is a legitimate business decision. Blocking a search crawler removes you from the answers. The trap is chronology: GPTBot appeared in August 2023, OAI-SearchBot arrived with OpenAI’s search work in 2024, Claude-SearchBot later still. So a rule written in 2023 either names GPTBot, in which case it is doing what you meant, or it is a blanket wildcard block that has silently caught every crawler invented since, including the ones that decide whether you get cited. Find out which you have.
Robots.txt is only the first gate. Your CDN or firewall is the second, and it is usually on without anyone having decided. Managed bot rules, rate limits and challenges return 403s to AI crawlers regardless of what robots.txt permits, and in our experience that layer blocks more of them than robots.txt does.
Which brings us to where you look, and this changed very recently. Google Search Console added generative AI performance reports on June 3, 2026, and as of August 31, 2026 they are available to all websites worldwide. They show impressions in AI Overviews, AI Mode and generative features in Discover, broken out by page, country, device and date (Google Search Central, June 3, 2026, checked August 31, 2026). Go and open yours. But note what they are: appearances on Google’s own surfaces. Impressions, not clicks, not queries, and not crawler visits. Whether OpenAI’s or Anthropic’s crawlers reached you at all appears in exactly one place, which is your raw server logs.
Can the engine resolve you to one entity?
An entity is not a brand. Your brand is what you say about yourself. Your entity is the single record the engine holds about you, with one description attached, and only one of those two appears in the answer.
This is the condition we failed, and it is worth showing the failure because it is the most common one and the least discussed.
We operate two domains. One is the consultancy. The other carries a book about running an operating business and trades as a registered name of the same company, so both are legitimately ours. The engine was forced to choose and took the one with more history and more corroboration. So it told buyers our AI visibility firm was an operations firm, sourced that from the book domain, and pointed them toward a competitor for the thing we actually sell. No amount of publishing would have fixed that, because the winning source was our own.
The symptoms are recognizable once you know them. The engine describes you as the wrong kind of business, names a competitor as the thing you probably meant, merges you with a similarly named company in another industry, or resolves your name to something unrelated, which is the constellation problem.
The fix is unglamorous: one name, one description, one set of links, repeated identically on your site, your LinkedIn, your Google Business Profile and every directory you appear in. Paraphrasing yourself across three places gives an engine three weak signals instead of one strong one.
Do independent sources agree?
Your own site is a claim, and claims are not worthless. One of our own domains beat every other source available for our name, which is the opposite of what most articles on this subject will tell you.
What outside sources buy you is agreement. When several describe you the same way, the engine has a consensus to state rather than a choice to make. When they do not, it picks, and it may pick the version of you that is out of date or wrong. That is why our own site being cited zero times matters less than the other half of the finding: no independent source describing us existed for the engines to weigh at all. There was nothing to agree with.
Corroboration is the slow part, and the part no monitoring tool gives you. Plenty of firms sell placement, listings and digital PR; none of them sell the outcome, because the sources that count are the ones you do not control. A LinkedIn presence describing the same entity in the same words, a verified Google Business Profile, a directory listing, a podcast appearance, a guest article on somebody else’s masthead. Each is a source the engine did not have to take your word for.
How do you check your own company in the next 30 minutes?
Do this before you buy anything or hire anyone. It will tell you which of the three conditions you are failing.
- Open a temporary or logged-out session in ChatGPT. Account history flatters the result, badly.
- Ask three questions: what is [your company], who are the best [your category] firms for [your customer], and how do I find a [your category] provider. Confirm the engine actually searched the web rather than answering from memory, because if it did not search the answer tells you nothing about your visibility.
- Record whether you were cited, mentioned, or absent, and record which sources it named instead. Who it names instead is the more useful half.
- Repeat in Perplexity, Google AI Mode and Copilot. Access varies by engine, by region and by week, and some will demand a sign-in or quietly degrade the answer without one. Where you cannot get a clean logged-out answer, use a fresh account with no history and write down that you did.
- Run each question twice wherever the engine lets you. These systems are not deterministic, and a single run is an anecdote, including ours.
- Open yoursite.com/robots.txt and search it for GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, Claude-User and Claude-SearchBot, and for any blanket rule that catches all of them.
- Write the date on all of it. Without a dated starting point you cannot prove any of this later.
The first run is usually uncomfortable. Ours was. It is still the most useful half hour available to you on this subject.
What should you not do first?
Do not buy a monitoring tool until you have run that baseline once by hand. Repeating this monthly is exactly what tools are for, and we use them. But until you have read one raw answer yourself, you will not know what a dashboard number is counting.
Do not fix your schema first if your name is split across two domains or two spellings. Valid structured data describing a confused entity just describes the confusion more precisely.
Be careful with anyone promising citations in thirty days. Ask what they are counting, on which prompts, on which engines, and under whose account. We have not seen a thirty-day claim survive those four questions.
And ask what any promised number actually means. We looked hard at whether we could honestly promise a mention rate and concluded we could not, so our agreement splits the two. We promise four things you can check yourself: that the AI crawlers can reach your site, shown in your own server records before and after; that your structured data passes Google’s own testing tool; that your business reads as one business across every listing we name; and a set number of published placements each quarter. Mention rate we report every month, alongside how much of the movement could be random noise, and we do not promise it. Anyone who hands you that number without the second half is selling you false confidence.
Why hire a firm the engines cannot find?
It is the obvious objection and it deserves a straight answer. What the audit above shows is the diagnosis we would run on your site, run on our own and published with the unflattring result intact. That is the entire content of the call we are offering. What it does not yet show is that we can fix it, because we are in the middle of fixing our own, in public and dated, including the parts that do not work.
The order is the piece we would defend hardest: reachable, then resolvable, then corroborated. Publishing into a site the engines cannot read is wasted effort, and corroborating an entity they cannot resolve just spreads the confusion further.
Authority compounds, and in our category a competitor has a head start we cannot simply buy back. What we can say is narrower: engines name several firms per query and revise constantly, so a head start is not the same as a closed answer. If you would rather buy from a firm with a finished case study, that is rational and we will tell you who to look at. If you would rather watch someone show their working on the problem you have, start here.
Get your free AI Snapshot: a written report in your inbox within one business day, yours whether or not you hire us.
Frequently asked questions
Why doesn’t ChatGPT know my company?
Usually because one of three conditions is missing. The engine has to be able to reach your site, which robots.txt, your CDN and your firewall decide between them. It has to resolve your company name to one clear entity rather than a duplicated or contradictory one. And it has to find independent sources describing you the way you describe yourself. Fail one and the engine either omits you or assembles a description from whatever sources it can reach, which is how companies end up described as something they are not.
Is my robots.txt blocking AI engines by accident?
It is common. AI companies run several crawlers with different jobs. OpenAI provides robots.txt control tokens for two of its four, GPTBot for training and OAI-SearchBot for ChatGPT’s search features, and says the two settings are independent; it notes that robots.txt may not apply to the user-initiated ChatGPT-User. Anthropic runs ClaudeBot, Claude-User and Claude-SearchBot and says all three honour robots.txt. A blanket rule written before these crawlers were split apart can block the search crawlers that decide whether you appear in answers. Your CDN or firewall can also return 403s to them regardless of what robots.txt says.
What is an entity, and why does it matter for AI visibility?
An entity is the single record an engine holds about your company, with one description attached. Your brand is what you say about yourself; your entity is what the machine concluded. If two domains, two company names or two descriptions of the same business exist, the engine picks one, and it tends to pick the one with more history and more corroboration rather than the one that is currently true.
How do I check whether AI engines can see my business?
Run three buyer questions through ChatGPT, Perplexity, Google AI Mode and Copilot, confirm the engine actually searched, and record whether you were cited, mentioned or absent along with who was named instead. Use a temporary or logged-out session where you can; several engines now demand a sign-in or degrade the answer without one, so where that happens use a fresh account with no history and note that you did. Repeat each question where the engine allows it. Then search your robots.txt for the AI crawler names, check whether your CDN is challenging them, and read your server logs. Date the results.
Can Google Analytics or Search Console show me AI crawler visits?
Not directly. Analytics measures human sessions. Search Console added generative AI performance reports on June 3, 2026, available to all websites worldwide as of August 31, 2026, showing impressions in AI Overviews, AI Mode and generative features in Discover, by page, country, device and date. Those report appearances on Google’s own surfaces rather than clicks, queries or crawler visits. Whether a third-party AI crawler such as GPTBot or ClaudeBot reached your site appears only in your raw server logs.
Do I have to let AI crawlers in to be cited?
Not always, but the exception is narrow. The New York Times blocks the major third-party AI crawlers in its robots.txt, including the search crawlers OAI-SearchBot and Claude-SearchBot, and ChatGPT and Claude still describe it accurately, because they can draw on decades of independent coverage about it. That is corroboration doing the work in place of access. Note that Google is a different case: the Times allows Googlebot, and Googlebot is what feeds AI Overviews and AI Mode. If your company is not already widely written about, third-party coverage will not carry you, and blocking the search crawlers removes the source that would.
