Why AI Search Names Your Competitor and Not You

One of the searches that lands on this page reads, word for word: “why is our b2b saas brand invisible when prospects search ai tools like chatgpt?” Note the “our.” That is someone asking on behalf of a company, probably after a prospect said they’d asked ChatGPT for options and got a shortlist without them on it.

It’s the right question. The answer starts with something most advice skips: “invisible in AI search” is at least three different failures, each with its own fix, and most brands spend their effort on the wrong one.

Watch what the model does before it answers

Open ChatGPT and ask the question your buyer asks, with the details a real buyer adds: team size, industry, the tool they’re replacing. Then watch the screen.

Sometimes it answers straight away. Sometimes it shows that it is searching the web, and the answer arrives with a list of sources. OpenAI built ChatGPT to decide by itself whether a question needs a web search. That fork decides everything after it, because the two paths draw on different material.

If it answered without searching, you’re looking at the model’s memory: what it absorbed about your category during training. If it searched, you’re looking at a retrieval pipeline. The model wrote its own searches, pulled pages from an index, read a handful of them and wrote from what it read. Your brand can be missing from either one, and the reasons have almost nothing in common.

Questions about current options, prices and alternatives give the model reason to check the web, so for B2B shortlists the retrieval path is where to start.

Step one: the model rewrites your buyer’s question

When the model searches, it doesn’t search for what your buyer typed. It breaks the question into several narrower searches and runs those. Google says so in plain terms: its documentation for AI Overviews and AI Mode explains that both may issue multiple related searches across subtopics before writing a response. Google calls this query fan-out. You can see the same effect in ChatGPT’s source list, where the cited pages often answer questions the buyer never typed.

This is the first place brands fall out. A buyer asks one question. The model asks five. Say the buyer wants an onboarding tool for a fintech with SOC 2 requirements. The searches that run might cover onboarding software for regulated companies, SOC 2 compliant vendors, pricing comparisons and alternatives to the category leader. Your homepage says “the modern onboarding platform.” It answers none of those. A comparison article on someone else’s site answers three.

So the first diagnostic question is not whether you rank for your keyword. It is whether you have the best available page for each narrower question the model asks on your buyer’s behalf.

Step two: the model can only read what its crawler could fetch

Next the searches hit an index. For Google’s AI features, that index is Google Search itself, and Google states that a page only needs to be indexed and eligible to show with a snippet to appear as a supporting link. For ChatGPT, OpenAI runs a crawler called OAI-SearchBot, and its crawler documentation says sites that opt out of it won’t be shown in ChatGPT search answers.

Two quiet mistakes live here.

The first is a robots.txt written in a hurry to block “AI bots.” OpenAI separates GPTBot, which gathers training data, from OAI-SearchBot, which powers search. A blanket block stops both. You opted out of training and, in the same line, out of being cited. Check the file, then your firewall rules, which can block unfamiliar user agents without anyone deciding to.

The second is JavaScript. When Vercel and MERJ tracked AI crawler traffic across Vercel’s network, the major AI crawlers downloaded JavaScript files but did not run them. Googlebot renders pages. Those crawlers read the raw HTML. If your pricing table, integration list or comparison page is assembled in the browser, much of the AI web receives an empty shell. The test takes a minute: view the page source, not the rendered page, and search for the sentence you want quoted. If it isn’t in the source, it doesn’t exist for them.

This is also where classic search work pays twice. A page Google hasn’t indexed can’t support an AI Overview, which makes the technical and on-page work that gets the right pages indexed the entry ticket to AI visibility rather than a separate discipline.

Step three: the model picks passages, and it prefers other people’s

Now the model has a pile of pages. It looks for passages that answer the narrow question it asked, then decides which to use and which to credit.

Two pieces of research describe what wins at this stage. The first is the GEO paper from Princeton, Georgia Tech and IIT Delhi, presented at KDD 2024. In their benchmark, adding citations, quotations from relevant sources and statistics to a page raised its visibility in generated answers by up to 40%. A benchmark isn’t your market, but the direction is clear. A passage that states a specific, checkable fact is easier for a model to use than a paragraph of positioning.

The second is less comfortable. Researchers at the University of Toronto compared AI search engines with Google across verticals and languages and found the AI engines leaned systematically toward earned media, meaning third-party, authoritative sources, over brand-owned and social content. Google’s mix was more balanced.

When a buyer asks for the best tools in your category, the model isn’t judging your product page. It is reading review sites, comparison articles, analyst roundups and trade press, and naming the products those pages name. If you are missing from those pages, you are missing from the answer, however good your own site is. We looked at the other half of this in what ChatGPT’s most-cited pages have in common: a large share of what it cites can’t be pitched at all, which makes the pages you can reach worth more.

Cited and named are two different results

This is the distinction most AI visibility reporting gets wrong. Being cited means your URL appears in the sources. Being named means your brand appears in the answer text. They move independently.

A brand can be named without being cited, because a third-party list the model cited mentioned it. A brand can be cited without being named, because the model lifted a definition from its blog and then recommended three competitors. For a buyer asking which tool to use, only the name in the recommendation does anything for pipeline. A report that counts citations of your blog can climb while you lose every shortlist.

So record both, and when you were named, record which page did the naming. That list is your real target list.

The memory layer moves slowly, and you can’t edit it

Back to the other fork: the answer that came without a search. It comes from training data, and no crawler setting or page rewrite will change it this quarter. The model learned what the web said about your category up to its cutoff. Your only lever is what the web says about you before the next model is trained.

The best public evidence on what shapes that is correlational. Ahrefs studied 75,000 brands against their appearances in Google’s AI Overviews and found the strongest relationships were off-site: branded web mentions (0.664), branded anchors (0.527) and branded search volume (0.392). It measured AI Overviews rather than model memory, and correlation isn’t cause, but it points the same way as the Toronto work: what others publish about you counts for more than what you publish about yourself.

Paid placement doesn’t get around this. Google places search and shopping ads inside AI Overviews, and Perplexity’s pause on advertising shows how unsettled the ad side of AI answers still is. Either way, a labeled ad beside the answer is not the model recommending you.

What you can change, in the order it pays

Once you see the pipeline, the fixes sort themselves by speed.

Access comes first because it takes hours. Allow OAI-SearchBot and the other search crawlers you care about, check the firewall, and make sure the pages you want quoted carry their content in the HTML. OpenAI says robots.txt changes take about a day to reach its search systems.

Your own pages come second, and they take weeks. Take the narrower questions the model asks and give each one a page or a section that answers it in the first two sentences under a heading that matches the question. Comparison pages, alternatives pages, plain pricing, integration specifics, compliance facts with the standards named: facts a model can lift and a buyer can check. This is where content built around the questions buyers ask before they buy earns its keep, and where a vague brand voice costs you.

Third-party presence comes third and takes months. Take the list of pages the engines cited when they named your competitors. Some are review platforms where you need a complete, current profile and real reviews. Some are comparison articles whose authors update them. Some are trade publications covering your category. This slow, unglamorous work is what moves the “best tools” answer.

Memory comes last because you don’t control the timing. The mentions you earn in the third step feed the next training run, whenever it happens.

How to run an LLM visibility audit yourself

You need a spreadsheet and an afternoon, not a platform.

1. Write down the questions your buyers ask before they buy, in their words. Pull them from sales call notes and support tickets, not your keyword list.
2. Run each one in ChatGPT, Perplexity, Gemini and Google’s AI Mode, in a clean session so past chats don’t shape the answer.
3. For each answer, record whether the engine searched, whether you were named, whether you were cited, and which competitors were named.
4. Collect every third-party URL cited in answers that named a competitor, and sort by how often each one appears.
5. Run the full set again a week later. One pass tells you little. A pattern across runs tells you a lot.

The useful output isn’t a score. It is two lists: the third-party pages that keep appearing, which is your outreach plan, and the questions where no page of yours was even a candidate, which is your content plan. If you want competitors mapped alongside and the gaps ranked, that is what our analysis of where your brand shows up online, and where it doesn’t covers.

Treat AI answers as a channel with its own rules, sitting beside the other channels you already run. Once you’ve watched the pipeline work, the lesson is plain. AI search holds no opinion of your brand. It has a reading list. Get onto the reading list, and the answer follows.

Frequently Asked Questions

How long does it take for a brand to start showing up in ChatGPT answers?

It depends which layer was broken. Access fixes register fast: OpenAI says robots.txt changes take about a day to reach ChatGPT search. New or rewritten pages have to be crawled and indexed before they can be retrieved, which usually means days to weeks. Getting onto the review sites and comparison articles the engines trust is outreach work measured in months. Training memory changes only when a new model is trained and released, on a schedule no outside company controls.

If our buyers use ChatGPT, does ranking on Google still matter?

Yes. Google’s AI Overviews and AI Mode take their supporting links from Google’s own index, so a page that isn’t indexed there can’t appear. The Ahrefs study also cites research from Seer Interactive that found a brand’s Google keyword rankings related more strongly to its ChatGPT mentions than link metrics did. What changes is the target. Rank for the narrower questions your buyer’s question breaks into when a model searches for it, not only for your head term.

Should we block AI crawlers to protect our content?

Separate training from search before deciding. OpenAI lets you block GPTBot, which gathers training data, while allowing OAI-SearchBot, which powers ChatGPT search. Blocking the search crawler means your pages won’t appear in ChatGPT search answers. One blanket rule can remove you from citation without anyone meaning to. Decide crawler by crawler, write the rules explicitly, then confirm your firewall or bot-protection service isn’t quietly overriding them.

Why does ChatGPT give a different list of tools every time I ask?

The model writes its own searches from your question, so small changes in wording change the searches and the pages retrieved. The same question can also produce different answers on different runs. The University of Toronto study found AI engines are sensitive to phrasing and differ from one another in the sources they draw on. That’s why a single screenshot proves very little. Run the same question set repeatedly, across engines, and look for the brands and sources that keep recurring.