AI search engine optimization (SEO/GEO/AEO) for manufacturers: how buyers find suppliers now

Your buyers now ask ChatGPT (and other AI search platforms) for supplier shortlists before they ever search Google. In a 2026 Semrush survey of 622 U.S. B2B professionals, with manufacturing the second-largest industry in the sample, 92% of the 519 who use AI at work said it shaped their vendor shortlist, and 41% said they start vendor research in an AI tool and turn to a search engine only to validate what it told them. Industrial buyers describe the same habit in their own words later in this piece.

To analyze what this change in buying behavior means for manufacturing organizations in particular, we audited 11 manufacturer and industrial company websites.

Every one of those sites was invisible for at least one buying query it should own.

This guide breaks down the findings from those audits, showing you how AI assistants choose which suppliers to recommend and the specific work that gets a manufacturer into those answers.

Read to the end, and you’ll have:

• The three gates your company has to clear before an AI assistant will shortlist your company
• The eight AI visibility defects that showed up over and over across our 11 audits, along with the fix for each
• The six fixes that appeared in almost every action plan, ordered by how quickly they pay off
• An honest measurement approach for a channel where the same question can return two different answers an hour apart

First, a real phone call we fielded that shows how buyer behavior has changed in a way a statistic never could.

A real phone call we got recently

The owner of an industrial equipment manufacturer called us after running an experiment. They weren’t happy with their current marketing agency (they had rated that agency a seven out of 10 in an internal discussion), so they decided to ask AI which agencies were worth looking at.

“With AI, I did a bunch of searches on ChatGPT, Gemini, Copilot, Perplexity. You know, stuff like that,” they told us.

Four AI assistants came back with a version of the same verdict: the incumbent agency was a generalist, and they needed a specialist.

The way the owner heard it: “You wouldn’t go to a general surgeon if you need heart surgery. You would go to a cardiovascular surgeon who does this stuff every day.”

A buyer with a budget asked an AI assistant who to hire and got a shortlist (which we were fortunately on).

Then our phone rang.

Notice how we weren’t involved in the research until the buyer had pretty much decided?

Strip the anecdote down and here’s what’s changed:

• A Google search used to hand your buyer a page of options: 10 blue links, a directory, a forum thread, you somewhere in the middle. The buyer compared, clicked around and built their own shortlist. You got to compete for a spot on it, and the buyer had to read quite a few of the ranking web pages to synthesize the precise answer to their question and decide what to do next.
• Now, an AI answer hands them the shortlist already built: two or three names, assembled from an industry directory, a “top 10” listicle, a forum thread and whoever’s site content answered the question best.

There is no page two.

They get a list of names, and a clear answer to their question or recommendation for how to solve their problem. If you’ve done the hard work to drive authority, trust and credibility for your company, you land on that list. And the comparison and elimination happen without you in the room.

Are your buyers really asking ChatGPT which suppliers to call?

In short, yes. And whether you make the list will affect whether your phone rings, or doesn’t.

Here’s the longer version:

• Forrester’s 2025 Buyers’ Journey Survey found 94% of business buyers used AI in their buying process, up from 89% the year before.
• Twice as many buyers as the year before named generative AI or conversational search a more meaningful source of information than any other source available to them.
• 61% now work inside a private AI tool their own organization provides, which means the conversation happens where your marketing cannot follow.

The Google side moved in step. SparkToro measured 68% of U.S. searches ending without a click in the first four months of 2026, up from 60% in 2024. Buyers are getting answers from the AI overview instead of links, and that answer names two or three companies rather than 10.

Our own call records say the same thing from the buyer’s side. We analyzed 32 sales calls with manufacturing leaders from a spring 2026 sample; 18 of them surfaced AI search, generative engine optimization, answer engine optimization or zero-click behavior without us raising it first.

We ask everyone who fills out a contact form on our site how they heard about us. ChatGPT, Claude, Copilot and other AI chatbots are becoming more and more common responses.

An executive at an industrial water-treatment company described the habit plainly: “Lots of companies right now are asking recommendations from the chatbots.” They then described buying that way themselves. “I want to hire a human resources company, and then I just ask Claude … what are the best options? And then I get a list of options. And those are the companies that I contact, you know?”

Traditional keyword volume data hides this completely. A components manufacturer we audited watched its core category terms register almost nothing in Google: roughly 20 searches a month for one term, zero for another. But buyers still ask those questions. They just ask them somewhere a keyword tool cannot see, and they get a more personalized response to their specific situation than a results page ever gave them.

None of which means buyers have handed over the decision. The head of an industrial-automation company put the boundary where most industrial buyers put it: “I’m not going to let AI diagnose me. I’m still going to go to a doctor. And I’m not going to let AI run my marketing. But I’ll let it point me in the right direction. And that’s honestly, that’s how I found Gorilla 76.”

AI narrows the field. People still choose.

How do AI assistants decide which suppliers to recommend?

An AI assistant builds its answer from pieces of evidence scattered across the internet. The system pulls short passages, usually a few hundred words at a time rather than whole pages, that sit close in meaning to the question. It reranks them, then writes an answer from what survives. (The exact passage size is a design choice each system makes and none of them publish it, so treat the number as a rough order of magnitude rather than a spec.) Your whole site never gets read. Your site either answers the question completely and succinctly, and gets lifted by the AI and cited in its response, or it loses to a competitor that answers the question better.

It helps to know that an assistant is working from two kinds of knowledge:

• What it learned in training. A broad snapshot of the web, refreshed only every so often, whenever the model’s maker retrains it. This is why an assistant can describe your company from memory. If the web said little about you when the snapshot was taken, the model barely knows you exist, and that stays true until the next refresh.
• What it looks up at question time. Buyer questions are usually too specific for memory: a spec, a lead time, a supplier for one application in one region. For those, the assistant runs a live search and reads pages the moment the question is asked. This in-the-moment research is called “grounding.”

Training decides whether the model recognizes your name. Grounding decides whether your pages get pulled into today’s answer. Most buying queries are specific enough to trigger the live lookup, which is why the work below matters more than whatever a model memorized about you last year.

Three things have to be true before your name comes out the other end. We call them the three gates:

1. AI has to be able to read your site.
2. AI has to be able to lift a clean answer.
3. AI has to trust your company as a source.

Fail any one of the three and you’re as good as invisible to the buyer looking for your product or service in the chatbot.

Let’s walk through each of these gates, one by one, and how to make sure your company clears each.

Gate 1: AI has to be able to read your site

The crawlers that feed AI answers have to get to your pages, and your answers have to live in HTML text rather than inside a PDF or a picture of a table. The reason is mechanical: HTML text is what a crawler actually reads, words sitting right in the page code, ready to lift. A PDF is a separate file many AI crawlers never open. And a picture of a table is just pixels; the numbers inside it don’t exist as text a machine can find, quote or compare.

This gate often fails quietly: Your site looks perfect to you and returns nothing to the AI chatbot trying to answer a buyer question.

What we check, and what fails it:

• Crawler access. Hard blocks in your CDN or firewall, and crawl-delay lines in robots.txt that throttle the crawlers to a trickle. Bing’s index is a primary source behind both ChatGPT search and Copilot; when Seer Interactive compared 500-plus ChatGPT citations against Bing across 100 queries, 87% of the citations matched Bing’s top organic results. Seer is careful to call that a match rather than proof of the plumbing, and OpenAI now supplements Bing with its own crawler. Either way, a rule that slows Bing slows ChatGPT’s view of you.
• Where your content actually lives. Spec sheets, case studies and certifications locked in PDFs are invisible to most AI retrieval. One company we audited had roughly 39,000 words sitting in PDFs against about 8,000 words in HTML.
• How the page renders. ChatGPT, Perplexity and Claude fetch pages; they don’t run them. Vercel measured real AI crawler traffic across its network and found none of the major AI crawlers render JavaScript: They will download your JavaScript files and never execute them, so anything your site assembles in the browser after the page loads does not exist as far as they’re concerned. Google’s Gemini is the exception, because it inherits Googlebot’s rendering.

What moves the needle: remove the blocks and delay directives, serve your money pages as plain server-rendered HTML and republish your best PDF-trapped proof as page text.

Gate 2: AI has to be able to lift a clean answer

Somewhere on your site there has to be a passage that answers the buyer’s question completely, in plain language, with specific numbers, without the system reading anything else.

This is where most good manufacturing sites lose. The expertise is there, but buried, and not easily extractable by an AI bot.

What we check, and what fails it:

• Does any page answer the buyer’s question directly? Capability overviews that gesture at everything (without saying anything at all) don’t give anything of substance that an AI can lift.
• Are the first 200 words the answer, or a warm-up? Make each word count, and make sure you answer questions clearly and succinctly without too much fluff.
• Are the specifics machine-readable? “Engineered for demanding applications” gives a model nothing. “Rated for 250 degrees Fahrenheit continuous duty” gives it a fact to quote and compare, and a reason to recommend you over someone else for a person’s very specific problem or application.
• Is the data in real HTML tables and lists, or trapped in a graphic a crawler can’t parse?

What moves the needle: one page per buying question, a direct answer in the first 40 to 60 words, spec tables as actual tables, your buyers’ own vocabulary in the headings and a specific citation for any claim you make. There is academic evidence this works at the page level: Adding statistics, quotations and citations to a page can raise its visibility in generative engine responses by up to 40%.

Gate 3: AI has to trust your company as a source

The rest of the web has to describe you consistently and say the same things about you that you say about yourself. This gate is the slowest to build and the hardest for a competitor to copy.

What we check, and what fails it:

• Entity consistency. One company name, one address, one description, everywhere: your site, your schema, directories, LinkedIn, data brokers. Contradictions read as unreliability.
• Third-party footprint. Independent pages that mention you: trade press, directories, reviews, forums. When every citation of your company traces back to your own site, one directory listing and LinkedIn, the model has nothing independent to trust.
• A findable named expert. A real person with a title, quotes in the press and a profile a model can retrieve.

What moves the needle: fix your directory and data-broker listings, standardize your own description everywhere it appears and put a named human expert in front of your expertise.

This third gate is also why working on your website alone is rarely enough. When we tally where a live AI answer actually got its raw material, the mix shifts by category, and the shift is worth knowing before you spend money.

• In some categories, third-party roundups supply more of the answer than anything else. We ran this tally on ourselves first. Across 25 live answer records from Claude, ChatGPT and Perplexity in July 2026, for the questions a manufacturer types when they’re shopping for a marketing agency, the “top 10” roundups came first, then review platforms and community discussion, then vendor diagnostic blogs. Every single time Gorilla 76 got named on a non-branded question, the source that put us there was somebody else’s list, never a page of ours.
• In other categories, the roundups barely exist. In one materials audit, the sources behind the answers were almost entirely manufacturer and distributor sites plus a light layer of trade directories. No forum or community content surfaced at all, which means the standard “get into the Reddit thread” advice would have been wasted effort in that category.
• Directories carry real weight either way. ThomasNet and its peers kept appearing across the industrial audits, and one company’s citations came back with a directory listing in every single run.

Two things follow. Most of the raw material in an AI answer lives on pages you don’t own. And which pages those are is a question to answer for your category rather than assume, because the answer decides where your off-site effort goes.

Also worth noting:

Appearing in an answer and being recommended in one are different outcomes, and the gap between them is wide. One analysis of 100 B2B “best of” queries in Google’s AI Overviews found that when a brand’s own self-promotional listicle got cited, the brand itself was left off the actual recommendation 69% of the time.

Which is why we track four states rather than one opaque, homogenous “AI search visibility” percentage:

Only the first one rings your phone. Recording all four is what lets you tell a near miss from a total blank, and the two problems have different fixes.

What’s usually broken on a manufacturer’s website?

We ran 11 AI-visibility audits for manufacturers and industrial companies in June and July 2026, and the same seven or eight things came up in almost every single one.

One company from that set shows what absence actually looks like. Decades in business. Its own testing lab. An engineer on staff who chairs an industry standards committee.

We put a handful of its buyers’ real questions to an AI assistant in a blind run. Blind means the assistant got the question and nothing else: no company name, no list of competitors, no hint about who we were checking on.

It was recommended first on one question, the niche application it’s best known for. On four of the remaining five it was never named at all, including its own core product category. Three of those four went to the same competitor, a company named consistently across dozens of independent distributor and directory pages while almost nothing on the open web describes our client.

Decades of proof, and on five of six questions the machine never said that company’s name.

The rest of the websites we audited failed the same way. Every company in the set was invisible for at least one core buying query it should have owned, and the defects clustered so tightly that the table below reads less like a survey and more like a checklist.

(We’ll detail how, exactly, to fix each of these defects later in this guide.)

If you aren’t listed when your buyers are searching for what you sell, someone is.

Is AI search different from SEO?

Mostly it’s the same work, with three real departures. Cyrus Shepard synthesized 54 experiments, patents and case studies on AI citation behavior and found that most citation factors are traditional SEO factors, sorting the 23 he scored under four signals: relevance, trust, topical authority and extractability. He is careful to say correlation is not causation, and so are we.

The cleanest way to see the difference is as two scoreboards:

1. The first asks: do we come up when someone searches? You win it with pages that match the keywords people type, building trusted links pointing at those pages and creating a site fast and clean enough to crawl. (The 7 core elements of an industrial marketing and sales strategy covers that work in full.)
2. The second asks: does the AI say our name out loud? You win it with an answer the model can lift straight off your page, a company described consistently everywhere it appears and trusted third-party sources talking about your company.

You can hold position 1 on the first scoreboard and be absent from the second. Most manufacturers have only ever watched the first.

Two examples from the bank of audits we’ve run for manufacturers:

1. One company ranks No. 1 to No. 4 on its core money terms in Google and does not appear at all in AI answers for its own differentiator query.
2. Another is the number 1 organic result for its brand name, gets a perfectly accurate branded answer from every assistant, and still went 0-for-6 on non-branded buying queries. Ranking is not the same as being recommended.

When the two scoreboards come apart, the failure is usually at retrieval. Here’s what that looked like in one audit, step by step:

• The model knew the company. Asked about the firm by name, every assistant described its services accurately, including the exact service in question.
• The model never surfaced it on buyer-phrased questions. When the question used the buyer’s words instead of the company’s words, the company vanished from the answers.
• The reason was vocabulary, and it’s mechanical. Retrieval pulls passages that sit close in meaning to the question as the buyer phrased it. The company’s pages described the service in internal language the buyers never use, or in thin, incomplete ways, so its passages never made the cut.
• The fix was structure and vocabulary on existing pages. A page whose heading asks the question the way buyers ask it, answered in the buyer’s own terms. No new channels, no net-new visibility tactics.

AEO, GEO and what the acronyms mean

Answer engine optimization (AEO) is the work of getting your pages retrieved and quoted by the tools that answer questions directly: ChatGPT, Perplexity, Copilot, Google’s AI Overviews. The name puts the weight on the answer. An answer engine doesn’t return a list of links for a buyer to work through; it returns a written response, and AEO is about making your page the one that response gets built from.

Generative engine optimization (GEO) describes the same work from the research side of the field. The term comes from the academic team behind the KDD paper cited earlier, who needed a name for optimizing content for engines that generate answers rather than rank pages.

In practice, marketers use AEO and GEO interchangeably.

Both differ from classic SEO mainly in what winning looks like. SEO earns your page a position in a list of links on Google or Bing. AEO and GEO earn your company a mention inside the AI answer itself, and the passage-level, entity-level work above is how.

But the overlap is large, and SEO still holds up fine as the umbrella term: A buyer asks a machine a question, whether that machine is Google or an LLM, and you want to be what the machine says back.

If your team keeps calling all of this SEO, they’re not wrong.

Worth knowing: marketers have settled on the vocabulary faster than buyers have. We mined our own sales-call corpus, 62 call records with manufacturing leaders from November 2025 through May 2026, and not one buyer originated any of those three terms. They say “AI,” “ChatGPT,” “SEO” and “chatbots.”

What do the manufacturers who do show up have in common?

Five things, and every clean win in our bank of audits had at least two of them:

1. A page built for one buying question. The wins traced to sharp application pages and to blog posts that answered the exact question a buyer asked, not to capability overviews trying to cover everything. Do this: pick the 10 buying questions you most want to own and give each one its own page.
2. Concrete specs and numbers in text a machine can read. An assistant quoted one manufacturer’s at-a-glance spec blocks almost word for word. One equipment manufacturer’s winning page opens with a real answer and a 3-to-24-month payback figure. Another’s FAQ pages name its competitors and real customer cities, and those pages anchored its lead placement on Perplexity. Do this: put a number on every claim, in an HTML table, directly on the page.
3. Write copy intended to be lifted by AI. One win came down to plain, concrete process descriptions that got quoted almost verbatim in all three runs we recorded. Do this: open each page with the direct answer in the first 40 to 60 words.
4. Owned media plus accurate entity data. One decades-old firm in the set gets an accurate profile from all four assistants, built on a recurring market report its industry actually reads, a long-running podcast and directory data that agrees with itself everywhere. Do this: publish something recurring your industry cites, and make every listing about you agree.
5. A named human expert. In one audit we mapped every competitor that beat our client on the buying queries it lost. All of them pair a recurring branded report with a named, press-quoted person behind it. The inverse shows up all over our audits: a standards-committee chair with no findable presence online, an owner unnamed on their own company’s site and a wrong CEO filled in by data brokers because there was no extractable leadership information to correct them. Do this: name your expert on your site, give them a real bio page and put them in front of the trade press.

No. 2 on that list is worth seeing at page level, because the gap is almost embarrassing once the two pages sit side by side.

Take this illustrative case of a generator manufacturer:

• A buyer asks which manufacturers make units quiet enough to install near occupied space, because noise is a hard constraint on the job.
• Seven competitors come back named. Every one of the seven has published an actual sound-level figure somewhere on its site.
• Our client builds the quieter unit — but never published the specific numbers that back that up on their site. The quiet-operation claim lives in adjectives on the website and in specifics only in sales conversations.

A machine can lift “rated at or below 55 decibels (dB) at 1 meter” and line it up against six other suppliers. It can do nothing at all with “exceptionally quiet.”

That’s a real finding from a real July 2026 measurement run (with the industry and product changed): seven competitors named on a spec-constrained question, every one of them publishing the number, our client publishing none of it and appearing nowhere in the answer.

None of the five mechanisms is something you can game, which is good news if you have real expertise:

• Keyword stuffing does nothing here; the model reads for meaning, not repetition.
• Bought backlinks don’t earn a mention by name.
• Mass-produced AI content adds pages that say what a thousand other pages already said.

What moves the answer:

1. Answering the question more completely than anyone else has.
2. Publishing the numbers and specifics only you have: tolerances, lead times, failure rates, real job outcomes.
3. Original research that third parties end up citing with your name attached.

Freshness helps, within limits. Ahrefs studied nearly 17 million cited URLs and found AI assistants cite content 25.7% fresher than organic results: 1,064 days old on average against 1,432, with ChatGPT preferring URLs 458 days newer than Google organic. That average cited page is still 2.9 years old, so age alone is not disqualifying. Like traditional search, AI assistants mostly cite content that has been around a while.

Redating a page without changing its substance does nothing.

What can a manufacturer do to show up in ChatGPT, Claude and Perplexity?

An executive at an industrial water-treatment company asked us this in almost these words: “So in fact, what can we do in order to appear in … those lists?”

Here’s a list of six things you can do, ordered roughly by how fast they pay off:

No. 4 is the one manufacturers underestimate. A buyer interviewed during one of the audits described their first move as typing the plainest possible word for the product itself and, in their words, “go simple.”

Your content, too, should “go simple.”

How would you know if it’s working?

Here’s the honest answer:

You can’t measure AI search the way your SEO dashboard measures rankings. Anyone who hands you a single tidy score is smoothing over how this actually works.

What you can do is triangulate with a handful of instruments, anchor on the one number that matters (qualified leads who found you through AI), and treat the qualitative signal from real buyers as data rather than anecdote.

Why your SEO dashboard habits won’t transfer

A rank tracker works because Google returns roughly the same results to roughly everyone.

AI answers don’t behave that way, for three reasons you can verify yourself right after you read this:

• The same question returns different answers run to run. Ask an assistant the same buying question twice in a row and you’ll usually get two different shortlists. SparkToro tested this at scale: 600 volunteers ran 12 prompts through ChatGPT, Claude and Google’s AI for 2,961 responses, and found under a 1-in-100 chance that two runs of the same prompt return the same list of brands. Order is worse; closer to 1-in-1,000. Marketers call this surface variance: The answer surface shifts under repeated measurement. One reading is a coin flip, and a trend line built from single readings is a coin-flip diary.
• Different assistants pull from different sources and behave differently. A score on ChatGPT is not comparable to a score on Gemini, because each assembles answers from a measurably different mix of pages. Ramp ran a 32-day experiment serving tracked content to AI bots and watched Claude surface its offer consistently while ChatGPT never surfaced it once, despite visiting the pages.
• The model can’t tell you why. When an assistant explains why it recommended someone, that explanation is not reliable evidence of how it decided. Behavior is the signal. The stated reasoning is color.

Our own measurement bank holds the sharpest version of that lesson. One buyer question, three surfaces, the same afternoon:

• Perplexity named the company and recommended it first.
• Google returned the company’s own page as the top organic result and produced no recommendation at all, because no AI Overview rendered for that query. Cited, not recommended.
• ChatGPT never mentioned the company once, across an answer running roughly 8,000 characters. A competitor took the top pick.

One question, one afternoon, three different verdicts on whether this company exists. Any single spot check would have told you something confidently wrong, and which wrong thing you got would have depended on which tab you happened to open.

A manufacturing leader we spoke with recently figured out the volatility problem without any help from us. “You can do the same search through them two times in a row, and one time your score out of 100 might be 10 and the next time it might be 32. So it’s not real consistent, but it gives you an idea, and I check that.” They apply the same skepticism to the answers themselves: “The thing about AI is it’s never going to tell you it doesn’t know the answer. It’ll give you an answer; it doesn’t mean it’s right. … So it’s like, okay, I got to verify what you told me is true.”

This is also why the “AI visibility score” a tool subscription sells you tends to feel wrong.

A single percentage averages all that variance into false precision, and you can’t take it apart to see which question, which platform, which day. We’ve watched a single-tool reading report 97% visibility against a reality that looked nothing like it.

A number you can’t decompose into individual query results is a number you can’t act on.

Why you can’t game it, and why that’s good news

Early SEO rewarded volume plays: target “what are generators,” publish a thin page, celebrate the traffic spike and never mind that nobody visiting that blog post was ever going to buy a generator. That playbook has no equivalent here, and we consider that a feature.

The model reads for meaning. It lifts the passage that answers the buyer’s specific question and ignores the one that repeats a keyword at it. The only reliable way to show up is to publish the most complete, most specific answer available, which happens to be the same thing that persuades the human who reads it after the AI quotes it. There’s no benchmark to game because there’s no single benchmark at all. The work is the ranking factor.

So we don’t celebrate traffic spikes, and we won’t promise you one. A manufacturer shortlisted by ChatGPT for one high-intent buying question is worth more than 10,000 visits from people asking what a generator is. (We hold our own content to the same standard, and we’ve published how we use AI in our marketing and content creation, since the bar applies to us too.)

What to measure instead

Since no single number is trustworthy, measure like a scientist with several imperfect instruments rather than a gamer with one score to beat:

The north star: a qualified lead who names AI

Every instrument above is a proxy except one. The point of this work is a real, qualified buyer saying some version of “I asked ChatGPT who does this and you came up.”

That’s the number to build your reporting around, and it takes deliberate instrumentation to capture:

• Put “How did you hear about us?” on every high-intent form, as a free-text field rather than a dropdown. This is where one manufacturer found “ChatGPT” typed in by hand. A dropdown without an AI option erases the signal.
• Ask again on the first sales call, and go one layer deeper: what did you search, which tool and what did it tell you about us? Log the answers somewhere your marketing team reads.
• Feed those answers back into the content plan. The questions buyers say they asked, in the words they used, are the next pages to build and the next prompts to add to the measurement basket. Treat that qualitative signal as the steering input for the whole program.

On timelines, the honest answer depends on which of the three gates you’re failing, and our 11 audits give the shape of it:

• The read gate moves fastest: unblocking crawlers and adding schema rated the smallest efforts in our action plans, usually a developer afternoon.
• The lift gate is content work, measured in weeks per page.
• The trust gate is the slow one: Directory cleanup, third-party mentions and a findable named expert build over months. Perhaps building and releasing a podcast (which is a slow burn). Perhaps an organic content strategy for the top leaders at your company.

Where does this data come from?

Every finding we describe as ours comes from 11 AI-visibility audits of manufacturers and industrial companies, run in June and July 2026, plus the monthly measurement runs that follow them and one audit we ran on ourselves. Where a number comes from somebody else’s research, it’s linked in the sentence that uses it.

The core of each audit is what we call a prompt basket, and it’s less exotic than it sounds. A prompt basket is a fixed list of questions, written the way your buyers actually phrase them, one question per buying intent. Five intents, five questions, put to four assistants: 20 answer records per cycle. Some baskets carry a sixth question where a client has a sixth intent worth watching.

Here’s each intent, with the kind of question it holds, using an industrial generator manufacturer as the example:

1. Researching a category. “What should I look for in an industrial generator for continuous duty at a food processing plant?”
2. Comparing suppliers. “Who are the leading industrial generator manufacturers for hospital backup power?”
3. Solving a problem. “Our plant’s backup generator keeps tripping under load. What causes that, and who can help?”
4. Vetting a shortlist. “We’re down to two generator manufacturers. What separates a good one from a great one for a critical-power project?”
5. Checking a specific company. “What does [generator manufacturer] make, and how do they compare to the bigger names?”

We put the same questions to ChatGPT, Claude, Perplexity and Gemini, and score every answer as one of the four appearance states: named and recommended, cited only, used without credit, absent. Then we re-run the identical basket monthly, so movement is comparable month to month rather than run to run.

Which questions go in the basket is decided by your sales-call and subject matter expert interview transcripts rather than by search volume, for the reason the near-zero-volume example makes concrete: the queries worth measuring often carry no measurable volume at all.

Qualitative input feeding a content plan is standard practice at any good content shop. What’s specific here is the combination of buyer language deciding the measured query set, an evidence gate on every audit claim and monthly re-measurement on an unchanged instrument.

Beyond the basket, each audit works through six questions about the site itself:

1. Can the crawlers that feed AI answers reach your pages? We test robots.txt rules, crawl-delay directives, and firewall and CDN challenges. We also fetch key pages the way ChatGPT, Claude, Perplexity and Bing do.
2. If they read your site, can they lift a clean answer? We score your key pages for direct answers, structure, tables versus images and how much of your proof lives in PDFs instead of page text.
3. Does your site tell machines what you sell? We read your structured data from the raw HTML: Organization, FAQPage and Product or Service markup, and whether it matches the visible page.
4. What do the assistants say about you right now? The blind basket runs, plus branded checks: ask each assistant about your company by name and grade the answer for accuracy.
5. Is your content built around your buyers’ questions? We map your pages against the questions from your own sales calls and look for the gaps, including vocabulary mismatches between what buyers say and what your pages say.
6. Does the rest of the web back you up? We trace where your citations come from, check your directory and data-broker listings for contradictions and look for a named, findable expert.

The data sources behind those six: a full crawl of your site, raw-HTML schema reads, live crawler-access tests, Google Search Console and analytics where access exists, keyword and authority data from Semrush, your directory and review-platform listings, plus the transcripts of real sales conversations that supply the buyer language.

What does a year of this work actually build?

A one-time audit is a snapshot: what the assistants say about you today, why, plus a ranked fix list. Useful, and incomplete, because the answers move.

The real system is a loop that runs monthly, and it helps to see a full cycle from the start.

The first month is the baseline. The audit above runs in full: what every assistant currently says about you across the prompt basket, which of the three gates you’re failing and where, and a fix list ranked by speed to payoff. This is the “you are here” map. Think of the year as a drive from Los Angeles to New York: the baseline hands you the map and the starting point, and every re-run after that is a road sign telling you to stay the course or take a better route.

Each month after that, the loop runs:

1. Fix and build. Work the fix list, and publish content engineered to answer specific buyer questions from the basket more completely than whatever the assistants currently cite.
2. Re-run the same basket. The identical questions, the same assistants, scored the same way.
3. Read the movement. Which questions moved from absent to cited or cited to recommended? Which didn’t move at all? Which competitor gained?
4. Fold what you learned into the next cycle. The questions that moved tell you what works for your buyers. The ones that didn’t tell you where the next month’s effort goes.

Every piece of content in that loop gets framed as a bet and written down before it ships: what we expect it to do, and what result would tell us to move on. A real one looks like this:

The system is built to find the winning ideas and tactics for YOUR company and YOUR context and YOUR buyers fast: more bets, each one cheap, each one scored against a written expectation. The bets that hit get doubled down on. And the ones that don’t still pay you back, because each one is a recorded answer about your buyers and your category that keeps anyone from spending money on the same question twice.

What compounds is the insight.

Month 12 starts from everything the first 11 months proved:

• Which questions your buyers actually ask assistants
• Which answers moved appearance states
• Which page structures got lifted word for word
• Which bets paid and which taught you something

A one-time audit hands you a to-do list. A year of scored cycles hands you a knowledge base about your own buyers that no tool subscription can generate, because three of its four inputs (your experts, your buyer conversations, your customers’ own words) live inside your company.

Frequently asked questions

Is this the same thing our SEO agency already does?

Mostly, with three real differences. Most AI-citation factors are traditional SEO factors, per Cyrus Shepard’s synthesis of 54 experiments, patents and case studies, with his correlation caveat attached. What’s new is passage-level extractability, entity consistency across sites you don’t own and a scoreboard of appearance states instead of ranks.

How long before we show up in AI answers?

It depends on which of the three gates you’re failing, because the fixes move at different speeds. Across our 11 audit action plans, crawler and schema fixes rated the smallest efforts (usually a developer afternoon), extractable content is weeks of work per page, and trust signals build over months. First citations of purpose-built pages might appear within one to two weeks and measurable pipeline impact might take three to four months.

Can anyone guarantee ChatGPT will recommend us?

No. The same prompt rarely returns the same list of companies twice, different assistants cite from different source mixes, and a model’s own explanation of its recommendation isn’t reliable. A guaranteed placement is a promise about a number nobody controls.

Do we need an llms.txt file?

No credible evidence supports it. In this study llms.txt scored dead last of the 23 citation factors ranked, and the author’s own note is blunt: He could not find any credible evidence or experiment showing llms.txt files influence AI citations at all. The work that moves appearance states is schema, extractable page text, entity consistency and a strong third-party footprint.

Does it matter if AI cites us but doesn’t recommend us?

It’s a partial state worth tracking rather than a win. One analysis of 100 B2B “best of” queries in Google’s AI Overviews found a brand’s own self-promotional listicle cited while the brand was left off the recommendation 69% of the time. The sharper finding sits underneath that number: The thin listicle where — surprise! — you rank yourself first doesn’t earn the recommendation. The engines pull the competitor names out of your own article and recommend them instead, and Google has been demoting exactly this kind of self-proclaimed-expert content. Third-party roundups carry weight because someone else wrote them. Your own version votes against you. That gap is why we record four appearance states instead of one.

Should we still care about Google?

Yes. The rise of AI search is not the death of Google search; buyers added a surface, they didn’t trade one for the other, and the two are connected. 68% of US searches now end without a click, per SparkToro, while AI Overviews and AI Mode run on Google’s own index and Search Console reports their impressions. Buyers also Google the names AI hands them before they call. Ranking well still isn’t enough on its own: one company we audited ranked 1 to 4 on its money terms and was absent from its own differentiator query.

Where this goes next

The owner of a niche equipment manufacturer described their position to us this way: “the stuff is so fledgling and it’s so new in our industry so antiquated that I don’t think very many of my competitors even know you can rank for AI let alone SEO so we’re still kind of top dog there.”

That advantage has a shelf life. But read it again from the other side: entire industrial categories are sitting wide open right now.

In our audits, whole buying questions had answers assembled from directories and generalists because no manufacturer had bothered to answer them. On one query, no manufacturer appeared at all.

For the companies that move first, here’s what seizing it looks like:

• Pick the 10 buying questions you want to own and publish the page that answers each one better than anything the assistants currently cite.
• Publish the numbers only you have. Your tolerances, lead times, payback periods and real job outcomes become the comparison baseline the machine quotes, and your competitors get measured against your figures.
• Put your expert’s name on it. While competitors stay faceless, your named person becomes the one the models and the trade press can actually cite.
• Re-ask the questions monthly and let the movement tell you where to press.

The buyer questions are already being asked. The pages that answer them are being written right now, by somebody.

Let that somebody be you.

If you want to know which questions your buyers are asking assistants, and what the answers currently say about you, that’s the work we do. The owner of an industrial repair contractor put the goal better than we could: “I want to be the answer.”