‘Why doesn’t my manufacturing company show up in ChatGPT answers?’

Does this sound familiar?

• You opened ChatGPT and asked it what you think your buyers might ask. Something like: “Who are the best suppliers of custom control panels for food processing plants?”
• ChatGPT gave a thoughtful answer, recommending three companies depending on the situation.
• And your company name was nowhere to be found in the response.

You have decades of proof, real certifications and customers who would vouch for you on the spot. And an AI assistant just confidently recommended a competitor you beat on quality every day of the week. But not you.

We ran AI-visibility audits on 11 manufacturers and industrial companies in June and July 2026, and every single company was invisible for at least one core buying question it should have owned.

The cause was almost always one of four things, ranked here by how often each one was actually true:

The top issue we saw is about what your website says — and the answers you provide within its copy — and you can start fixing that without touching a line of code.

The rest of this guide breaks down why, exactly, you might be absent from ChatGPT (or Perplexity or Gemini or Claude) … and, more importantly, what you can do about it.

By the end, you’ll have:

• The four causes, in the order to check them, with how often we actually see each one
• A check you can run yourself without hiring anyone
• The bot names and settings that matter for ChatGPT, Claude, Perplexity, Gemini and Copilot, taken from each company’s own documentation
• An honest read on which of these you can fix in house, and which you can’t

‘But I rank in Google! Shouldn’t that mean I should show up in ChatGPT?!’

No, and this is the part that catches everyone. Google ranks pages. AI assistants fetch a handful of passages, decide which ones answer the user’s exact question and then decide what to do with what they found: recommend you, cite your page without recommending you … or leave you out entirely.

Your position on a Google results page is not a clean input to that process.

One company from our audit set shows how far apart the two can sit. Its site ranks No. 1 to No. 4 on its core money terms in Google, and it does not appear at all in AI answers for the questions that play to its real differentiators.

If your site lacks educational content built for AI search, and you haven’t worked to build trust signals from third-party sources, your company will be effectively invisible when a buyer asks an AI assistant for what you sell.

One sales leader we talked to put the stakes more bluntly than we ever would: You’re either going to appear, or you’re not.

Cause 1: Your website isn’t answering the questions your buyers actually ask

All 11 companies we audited had a problem here, and it has nothing to do with the quality of your team’s expertise. It has to do with whether that expertise is captured on your website in a form a machine can lift.

Start with how buyers actually phrase things. The owner of one company we work with described his own buyer’s moment of need this way: “Somebody goes to ChatGPT or some other service and they say, ‘I’ve just had a fire. And I need to replace some rafters. Is there an engineering firm I can call? Is there anybody that does this?’”

That buyer never typed a product category. They typed a problem.

An AI assistant hunts for a passage that answers that exact problem, grabs it and moves on. So your answer has to sit directly under the question, in plain language, with real numbers in it. And it has to sit in text and HTML. An AI bot reading your page sees nothing inside an image of a spec table, and nothing inside a PDF it never opened.

We saw it repeated in the 11 audits we did:

• Roughly 39,000 words locked in PDFs against about 8,000 words in HTML, at one company
• 86.2% of pages with no subheads at all, at another
• A 2,694-word FAQ page with no subheads, so the whole thing read as one undifferentiated block
• Application pages of 200 thin words, published by companies whose engineers can quote tolerances from memory

You can be the best option in real life, and still be invisible when a buyer asks AI

Take a manufacturer with decades in the business, its own accredited test lab and an engineer on staff who chairs an industry standards committee.

By any measure, these are the experts.

But when we put a handful of its buyers’ real questions to an AI assistant in a blind run, their name was nowhere to be found. Blind means the assistant got the question and nothing else: no company name, no competitor list, no hint about who we were checking on.

It was recommended first on one question, the niche application it is best known for. On four of the remaining five it was never named at all, including its own core product category.

See the full breakdown in the chart below:

What you can do right now

• Structure your content, so it puts the answer in the first two sentences under the question. If a reader has to scroll to find out whether you do the thing, so does a machine.
• Write subheads as the questions your buyers ask, then answer each one completely enough to stand alone.
• Move real numbers into page text. Tolerances, ranges, lead times, capacities, test results. A specific number is the single most liftable thing on a page.
• Lift specs out of PDFs into HTML, and ship tables as real HTML tables, never as images. Keep the PDF as the printable version.
• Publish the proof you already have internally. Six of the 11 action plans included this same recommendation. One company had an unpublished 40% efficiency gain a customer measured after implementing its solution. Another had a story about roughly $150,000 in tooling costs a customer avoided. Both sat in a drawer — not on the site.

Cause 2: Nobody besides you vouches for your company

AI assistants like ChatGPT, Claude and Perplexity use third-party trust signals — review sites, trade publication coverage, directories, social channels — to decide who to recommend. When an assistant recommends a supplier, it is staking its answer on that supplier, so it leans on what it can corroborate from more than one place.

Your own website is one source. But any company can say anything about how great they are.

Answering real buyer questions on your site gets you considered. But third-party corroboration is what gets you confidently recommended, and it lives mostly on property you don’t own: directories, trade publications, distributor sites, review platforms, LinkedIn, standards bodies. It is the slowest of the four causes to move and the hardest for a competitor to copy, which is exactly why it’s worth building.

What the audits found:

• All 11 had a third-party footprint too thin to support a recommendation. At one, every citation we could trace led back to its own website, one directory listing and its LinkedIn page.
• Directories fill the void. ThomasNet appeared in every single run for one client’s category. If nobody vouches for you, the assistants fall back to whoever’s listed there.
• One rival’s distributor network formed what the audit called a “citation moat”: dozens of independent pages describing the same product the same way.
• The competitor that won across every query paired a recurring branded report with a named human analyst — which earned it real coverage in the press.

The mirror image showed up across the whole audit set: real experts nobody outside the company can find. A standards-committee chair with no author page. An owner unnamed on his own website. One company had no extractable leadership information on its site, so when we asked Claude about it by name, Claude fell back on data brokers and named a wrong CEO.

What you can do right now:

• Claim and correct the directories your category actually gets cited from. Find them by asking ChatGPT or Perplexity a category question and reading the sources it lists back.
• Get your named experts onto the record. An author page with credentials, a byline on your technical content and quotes in the publications your buyers read. Perplexity weights named authors with credentials heavily.
• Build a review corpus where your buyers look for one. One contractor we audited had zero reviews while a direct rival held more than 30 on Google.
• Publish a number nobody else owns. The biggest recurring opportunity in the whole audit set, and the least used: Five of the 11 action plans recommended publishing an owned category statistic, because nobody in the category had one. One of those companies runs a lab producing 60,000 data points a year and has never published a benchmark. A statistic only you can supply gets cited by everyone writing about your category.

Cause 3: The web doesn’t describe your company the same way everywhere

Nine of the 11 companies published facts about themselves that contradicted each other. Different names, different founding years, different service lists, different addresses.

This is not an inconsequential issue. In three of the 11 we watched an assistant repeat or act on the bad data live:

• Claude, unable to confirm one company’s leadership from its own site, named a wrong CEO pulled from data brokers.
• One company launched a new product under a name two competitors already used. Ask Claude or Gemini about it and both split the name in half: they credit the company for part of it, then hand the product itself to the competitors who got to the name first.
• A third was confused with a similarly named business in the same state.

An AI assistant resolving who you are has to pick between two versions of you. Every conflict it finds is a reason to describe you vaguely, describe you wrong — or ignore you entirely, and reach for a competitor it can pin down.

Brian Dean, who spent the past year testing AI visibility on his own site, landed on the same rule from the practitioner’s side: if your homepage says one thing and your LinkedIn says another, “the model gets confused.” Every page, profile and mention needs to reinforce the same claim about who you are and what you’re best at.

Structured data tells the same story. All 11 were missing the markup that tells an assistant what they sell. One had none at all. Another carried more than 1,700 structured-data statements without a single Organization, Product, FAQ or Article among them, just a CMS filling in boilerplate rather than anyone describing the business.

What you can do right now:

• Pick one form of your company name and one canonical description, then use them everywhere. We ran this audit on our own site too, and found 711 instances of our name written without the space in it (“Gorilla76” instead of the correct “Gorilla 76”), plus a page misspelling one of our co-founders (“John Franco” instead of the correct “Jon Franko”). Google’s AI answers had started repeating the misspelling back to us — and we’ve been working on a data cleanup of our own since that audit.
• Audit your own properties first, then work outwardly with directory listings, data-broker profiles, distributors’ descriptions of your products, your team’s LinkedIn pages.
• Retire or redirect legacy websites. One company we audited had two old sites still live, last touched in 2016 and 2020, quietly contradicting the current one.
• Fix your structured data. It’s the cheapest item on this list and it appeared in all 11 action plans: Organization and FAQ markup, plus Product or Service markup, with real values rather than CMS defaults, and “sameAs” links connecting your LinkedIn, directory listings and any Wikidata entry.

Cause 4: Something is blocking AI from reaching your site

We found crawl-side problems at eight of the 11 companies. Three were hard blocks. Five were a crawl-delay line in robots.txt, which blocks nothing outright and quietly starves the crawler anyway. That matters because crawl-delay throttles Bing’s crawler, and Bing feeds retrieval for both ChatGPT and Copilot.

A block hides in three places, and only the first is where people look:

1. Your robots.txt file
2. A bot rule at your CDN or firewall that nobody on the marketing side has ever seen
3. That crawl-delay directive

At one company we tested, requests carrying the ClaudeBot name came back 403 Forbidden on every URL including the homepage, three times out of three, while GPTBot, ChatGPT-User, OAI-SearchBot, PerplexityBot and Google-Extended all returned 200 OK on the same pages. Nothing in that site’s robots.txt mentioned any of them. The rule sat at the CDN edge, where no robots.txt review would ever find it.

What you can do right now: check for a block yourself

Every crawler announces itself with a name (its user agent), and a bot rule fires on that name. Borrow the name, and you see exactly what the bot sees.

With a free Screaming Frog license, go to Configuration, then User-Agent, and set the custom user agent to GPTBot. Crawl your homepage and a few key product pages, and read the status codes. Repeat with ClaudeBot, OAI-SearchBot, PerplexityBot and Bingbot.

One crawler refused while the others pass means a rule aimed at that crawler. All refused means a blanket bot setting at your CDN. Either way, the fix is one conversation with whoever manages your website or CDN account.

Does fixing a block put you back in the answers?

Usually not on its own, and the company with that 403 is the clearest example we have.

The blocked bot, ClaudeBot, builds Anthropic’s stored copy of the web. A separate fetcher reads pages live when a user asks a question, and the edge rule didn’t touch it. So during that same audit, Claude’s live web search reached this company’s site and cited five of its own pages.

The block was real and worth closing. But on every question phrased the way a buyer with a problem actually asks it (worker injury, throughput, durability in a specific application), the company was absent because the pages that would have answered those questions did not exist.

Close the block. Then go back to the top of this list, because the first two causes are almost certainly waiting for you there.

Which of these is most likely your problem?

Check them in the order below. It’s deliberately different from the likelihood ranking at the top of this page: The crawler check goes first because it takes minutes, and the answer gap sits second because nothing downstream matters until a liftable answer exists.

How do you check this yourself?

1. Write down five questions your buyers actually ask, in their words. This is where companies lose most often. One fabricator we audited had buyers who ask about PPAP documentation and ISO 9001 certification before anything else. Its quality page never said “ISO 9001” in text, and the certificate was a PDF sitting on a retired domain. It was absent in every run on those supplier questions, and a competitor whose page spelled out its PPAP process in plain text won every one of them.
2. Ask each question blind. Fresh chat, signed out and never name your own company. The moment you name yourself, you are testing whether the model has heard of you, which is a much easier test than whether it recommends you to a buyer who has never heard of you.
3. Record one of four outcomes per question, and write down who got named instead. The four outcomes cost you different things. See the chart below.
4. Test your crawler access using the user-agent check above, or hand that section to your developer.

Absent on any of them?

Walk that one question through the causes in check order, and stop at the first one that fails. Many defects were near-universal across our audit, so a list of things wrong with your website is a list of things wrong with every website. What makes a defect yours is that it sits between one specific unanswered buyer question and the answer.

Why does ChatGPT give you a different answer every time you ask?

Because these systems are probabilistic, and the variation is wider than almost anyone expects.

SparkToro measured this at scale: 600 volunteers ran 12 prompts through ChatGPT, Claude and Google’s AI Overview a combined 2,961 times. Their finding, in their words: “there’s a <1 in 100 chance that ChatGPT or Google’s AI, if asked 100 times, will give you the same list of brands in any two responses.”

Four things move your answer:

• The run. Same question, same assistant, five minutes later, different shortlist.
• The assistant. ChatGPT, Claude, Perplexity and Gemini all assemble answers from a measurably different mix of pages.
• The surface. ChatGPT logged out on a free account and ChatGPT logged into a paid workspace cite differently. Some sessions search the web live, some answer from memory.
• Your own history. The trap, because it bends the result the way you want it to bend. If you’ve ever asked about your own company before, the assistant has context a buyer would not.

Here’s what that looks like on real records. One buyer question, one company, four surfaces, the same afternoon in July 2026:

• Perplexity, signed in on our own account: recommended the company first. We threw this reading out. The account had asked this question before and memory was on, so it shows personalization, not what a buyer sees.
• Gemini, fresh session, signed out: recommended the company first. Clean read.
• Claude, blind run with live web search: named the company second, and cited one of its own guides as a source.
• ChatGPT, signed in on our agency workspace: never mentioned the company once, across an answer running 8,094 characters. A competitor took the top pick.

Three verdicts plus one discarded reading, same question, same day. And the discarded one was the flattering one. The check you run on your own logged-in account is the check most likely to tell you what you want to hear.

One manufacturing leader we work with found this on his own, watching his score on a free visibility tool swing from 10 out of 100 to 32 between back-to-back runs.

It’s our stance that any single score — “85% AI search visibility!” — flattens infinite variance much too cleanly.

So how do you actually measure this? Build a prompt basket.

A prompt basket is a fixed list of the questions your buyers actually ask, written down once, then asked of every assistant on a schedule.

Any single reading is close to a coin flip — but a fixed set of questions re-run on a schedule is what turns coin flips into a directional measurement.

Four things go into building a prompt basket:

• Intents, not phrasings. What the buyer is trying to accomplish? Are they trying to find a supplier? Compare options? Solve a problem? Check whether you’re qualified? Pick four to six.
• Several phrasings per intent. In SparkToro’s follow-up test, human-written prompts aimed at the same thing averaged a semantic similarity of 0.081, which is close to no overlap at all. Test one phrasing and you’ve measured that phrasing; test several, and you’re getting closer to measuring your visibility on that buyer intent.
• Every surface, signed out wherever the platform allows it.
• Repeat runs on the questions that matter most. Coverage tells you where you stand. Repeat runs are what let you say you own a question.

Each cell of that grid records the outcome, who got named instead and which sources were cited. What you get out is a per-question map of where you’re absent, who’s standing in your spot and which pages the assistants read to decide. That maps directly onto what to write next.

Is this different in Perplexity, Gemini, Claude and Copilot?

Yes, in one way that catches almost everyone. Every one of these companies runs several bots, and each bot does a different job. Think of it as three separate doors into your website.

• The library door. A bot that builds the company’s stored copy of the web, which is what the model learns from.
• The index door. A bot that builds the search index the assistant looks things up in.
• The errand door. A fetcher that reads a specific page in the moment a user asks about it.

They’re controlled separately. So “block the AI bots” usually closes one door and leaves the others open, or closes a door you needed open. Blocking the bot that collects training data does not remove you from the assistant’s answers. Blocking the bot that builds the search index can.

Somebody at your company may have already made this decision, possibly years ago, possibly to protect your content from AI training. That was a reasonable decision. It may also be costing you answers, and nobody has checked.

Ask whoever manages your website: which AI bots are allowed, where that rule lives (robots.txt, the CDN or both) and when it was last reviewed. Then check it against this table.

Sources: OpenAI, Anthropic, Perplexity and Google publish this themselves.

What can you do yourself, and where do you need help?

More than most agencies will tell you. Everything below on the left costs you nothing but time, and this piece gave you the instructions for all of it.

Take a look at the chart below, outlining the work it requires to move the needle for each cause of AI search invisibility:

Notice there’s nothing glamorous here.

This is hard, sustained work, and that’s exactly why it moves the needle: your competitors mostly won’t do it, or won’t do it well.

Expect meaningful movement around month three or four, compounding after — only if you’re doing regular, sustained work to tackle the biggest technical, content and trust gaps you identified for your site.

If you need help along the way, we’ll meet you where you are.

And if you want to run this in house, we offer consulting: we audit your site, hand you the findings with the receipts behind each one — and guide your team through the order to fix things in.

Frequently asked questions

We rank number one on Google. Doesn’t that mean AI will find us?

No. One company in our audit set holds position 1 to 4 on its core money terms and is absent from AI answers for its own differentiator query.

Are we too small for AI to notice us?

Size isn’t the gate. In our audits, whole industrial categories were sitting unclaimed, and on one buying question no manufacturer came back by name at all. Being early in a category nobody has answered is an advantage.

Is our robots.txt file the whole story?

No. Two of the three hard blocks we found sat at the CDN or firewall layer, where reviewing robots.txt cannot see them. Test the crawlers directly using the user-agent check in this guide.

If we block AI training, are we also blocking AI search?

Not necessarily. OpenAI, Anthropic and Google all run separate crawlers for training and for search, with separate controls. The catch is crawlers that do both jobs, which some settings handle as a single decision.

Can we do this ourselves or do we need help?

You can get to real insight on your own. The five-question check, reading your robots.txt and putting a real specification number on a page you already own need nobody. The per-crawler test needs a developer or an afternoon with a free crawling tool. But getting recommended and cited takes sustained work — by you or an outside party.

How do we know an AI-visibility number is real?

Ask the same questions repeatedly, across assistants, on a schedule, signed out where you can. A single reading tells you almost nothing. And if a tool hands you one homogenous “AI visibility” score, it’s compressing a highly variable system into far too simple a number. The truth is more complex.

Where to start

Run the five-question check today. It costs nothing and tells you more than any tool subscription.

Then decide how you want to feel about what comes back. One executive we work with watched our audit find his brand in 2.6% of the AI answers we tracked for his category. The exact number matters less than what it told him. His buyers were asking, and he wasn’t in the answers. It stuck with him. He created an internal project on the spot: “97% or bust.”

That’s the right response — to take action — even if optimizing for a homogenous skill oversimplifies the problem. In most industrial categories, almost nobody has done this work yet. The company that publishes real answers and earns real corroboration first gets to be the name in the shortlist while its competitors are still arguing about attribution.

AI search also doesn’t live alone: It sits inside the same strategy work as the rest of your marketing, and the 7 core elements of an industrial marketing strategy is the map for that bigger picture.

Your buyers are asking questions about your products and your services in AI search right now.

ChatGPT is going to recommend somebody. Don’t let it be your competitors.