Only 1.6% of the independent US manufacturers we audited have a real FAQ schema an AI assistant can extract with confidence, and three in four have no detectable FAQ content at all

At Sell with Marketing we ran a live audit of 200 independent, privately held US manufacturers (most of them family-owned), to measure how prepared they are for a shift that is already underway. In our own work with B2B clients (companies that sell to other businesses, not to end consumers) we see it more and more: buyers who show up to the first call after already asking ChatGPT, Perplexity, or Claude who can make what they need, before any human contact.

The discipline that measures this preparation is called AEO, Answer Engine Optimization: optimization for AI answer engines. (Internationally, "AEO" also refers to Authorized Economic Operator, a customs compliance certification: a different topic entirely, unrelated to this study.) This is what we found, with a public, reproducible methodology. We did not find a comparable public study for independent US manufacturers; if one exists, tell us and we will cite it.

Executive summary

  • We audited 200 independent US manufacturers across dozens of sub-sectors. 184 returned a working homepage (HTTP 200) and are the base for every percentage in this study, unless noted otherwise.
  • 1.6% (3 of 184) have real FAQPage structured data, the format an AI model can extract with the most confidence. A further 21.7% (40 of 184) have a detectable FAQ section identified by heading or link text, without the structured markup behind it. Combined, 23.4% (43 of 184) show some form of detectable FAQ content.
  • 8.2% (15 of 184) have none of the four main signals we measured (FAQ content, structured data, an llms.txt file, or a meta description).
  • 72.3% (133 of 184) have some structured data, mostly generic.
  • 35.9% (66 of 184) have a real, verified llms.txt file.
  • 2 of 182 sites (1.1%) with a readable robots.txt block at least one of the crawlers that generate live AI answers.
  • We measured the signals an AI model needs to be able to cite you. Whether a brand actually shows up inside a ChatGPT or Perplexity answer is a different thing: it is not fixed or predictable, it changes with every question, and it was not part of this study.

Check your own site in three minutes

Before the full findings, three things you can check yourself right now, on your own site:

  1. Open yourdomain.com/robots.txt in a browser. Look for two things: a Disallow: / under User-agent: *, which blocks every crawler at once, and a Disallow: / under any of PerplexityBot, OAI-SearchBot, ChatGPT-User, Claude-User, or Claude-SearchBot. If neither is there, these crawlers are not blocked at the robots.txt layer. That is not the whole picture: a firewall or CDN can still block them underneath, and robots.txt will not show it.
  2. Right-click your homepage, "View Page Source," and search for application/ld+json. If nothing comes up, an AI model reading your raw HTML sees no structured data about your company, regardless of what a human visitor sees.
  3. Disable JavaScript in your browser (or use a text-only browser extension) and reload your homepage. If the page goes blank or loses its main content, that is close to what most AI crawlers see, since the majority do not execute JavaScript.

Why this matters now

For decades, an independent manufacturer's competitiveness was measured on the shop floor: production capacity, certifications, lead times, price. Those variables still matter, but they are no longer the first thing an international buyer evaluates.

A buyer who is diversifying suppliers, reshoring part of a supply chain, or simply looking for a new domestic manufacturer increasingly starts that search by asking an AI assistant directly: who makes this, where, at what capacity. Unlike a traditional search engine, which returns a list of links for the user to sort through, these tools select and cite the sources they consider clear and verifiable enough, directly. What they cannot read and cite does not make it into the answer.

That turns the structure of a manufacturer's website into a filter that operates before any negotiation, before any quote, before any human contact happens at all. It is not just ChatGPT or Perplexity, either: Google itself already answers a growing share of searches directly with AI Overviews, the AI-generated summary Google places above the results, before a user clicks anything. The same mechanism this study measures for AI assistants increasingly applies inside the search engine a manufacturer already uses to track its own organic traffic.

What AEO is, and why it is different from SEO

SEO (Search Engine Optimization) is the discipline that, for 25 years, decided whether a business showed up on the first page of Google. AEO does not replace it, but it measures something different: whether a site's content is written and structured so an AI model can read it, treat it as a reliable source, and use it to build a direct answer for someone asking about that business. This is sometimes also called GEO (Generative Engine Optimization); it is largely the same idea under a different name. At Sell with Marketing we treat them as one job: not SEO versus AEO, but SEO and AEO together.

A site can rank well on Google and still be practically invisible to an AI answer engine, if its content is not organized in a format the machine can extract: clear questions and answers, verifiable data about the organization, specific information about products and capabilities, instead of only generic marketing copy.

How we did it

This section explains our method in the technical detail needed for anyone to verify it, or for your marketing or IT team to review it. If you would rather skip to the results, jump to "Finding 1" below.

We ran an automated audit script, written in Python, built specifically for this study following our own methodology, Argos v1.1 (module 05, AEO). The principle we followed throughout is that no data point is estimated, inferred, or filled in: whatever we could not measure live, we state explicitly, with the reason why.

Sample and how it was built: 200 independent, privately held US manufacturers across dozens of sub-sectors (furniture, textiles, food and beverage, agriculture, hardware, packaging, metal fabrication, and more). The candidate list was assembled through a directed web and AI-assisted search for independent US manufacturers, then screened company by company against the ownership evidence described below. We note the obvious limitation: a list built this way is more likely to surface companies that already have some findable web presence, which would tend to bias the results in this study upward, not downward. Each brand in our source list then carries a documented ownership note (public evidence such as revenue figures or LinkedIn company data confirming privately-held, independent status) that we used to screen out subsidiaries of larger conglomerates before the audit ran; 169 of the 200 (84.5%) carry explicit evidence of family or multi-generational ownership specifically, and we describe the full sample as "most of them family-owned" on that basis, not as a claim about all 200. This is a purposive sample assembled by our research process, not a random sample and not a census of US manufacturing.

Measurement date: August 23, 2026. This is a snapshot of that day; sites change, and auditing the same domain again in a few months could produce a different result.

How we identified ourselves: every request identified our crawler honestly as SWM-ARGOS-Research-Bot/1.0, with our contact information visible in the user agent. We never impersonated another bot or a human browser, and we honored each site's own robots.txt Disallow rules, even when that meant we could not measure it.

Limits we set ourselves: a maximum of 5 concurrent domains, with a random 2 to 3 second pause between requests to the same domain.

What we measured on each site:

  • robots.txt: the rules declared for 13 AI user agents, split into two families: live-answer crawlers (OAI-SearchBot, ChatGPT-User, Claude-User, Claude-SearchBot, PerplexityBot, Perplexity-User) and training crawlers (GPTBot, ClaudeBot, Google-Extended, Applebot-Extended, CCBot, meta-externalagent, Bytespider).
  • llms.txt: we verified real existence (HTTP 200 with Content-Type: text/plain or text/markdown), ruling out "soft 404s".
  • sitemap.xml: we verified existence.
  • Structured data: we extracted every JSON-LD block on the homepage and its types, recursively.
  • FAQ-style structure, measured at two confidence levels: (a) real FAQPage markup declared in JSON-LD, the format an AI model can extract with the most confidence, and (b) a lower-confidence heuristic: "frequently asked questions" or "FAQ" text found in a heading or a link on the homepage, which signals a human-visible FAQ section but not necessarily one an AI model can parse into distinct question/answer pairs.
  • Meta description: presence and length.
  • H1 and H2 headings: counted in the page's raw code.

What each signal actually measures, and its honest limit: we measured the code (HTML) the server delivers, which is the same thing most AI answer engines read, since the majority do not execute JavaScript. A site built as a single-page application, which injects its content with JavaScript after the page loads, can show a human browser more signals than it shows an AI model. That gap is not a measurement error on our part: it is exactly the kind of finding this study is meant to expose, because it is the same limitation any answer engine runs into when it reads that site.

Base for our percentages: of the 200 brands, 184 returned HTTP 200 on their homepage on the day of the run, and are the base for most percentages in this report. The remaining 16 are excluded from every percentage: 8 due to a connection error or timeout, and 8 that responded with an HTTP error code (six with 403 Forbidden, one with 429 Too Many Requests, one with 421 Misdirected Request). Rather than name individual companies next to an error code our own audit traffic may itself have triggered, we are publishing the full list of 200 audited domains to any journalist, researcher, or brand on the list who wants to verify or reproduce this run directly: write to hello@sellwithmarketing.com. We did not attempt to characterize the robots.txt rules of those 16 excluded sites. Of the 184 sites with a working homepage, 182 had a robots.txt we could read, and none of those 182 blocked our own crawler. The other 2 had a robots.txt request that failed independently of their homepage status, so we have no verifiable data on their crawl rules. That is why the AI-crawler data in Finding 1 below has a base of 182.

Finding 1: access is mostly open, but not universal

Result: none of the 182 sites with a readable robots.txt block all live-answer crawlers, but 2 distinct sites (1.1% of 182) block at least one live-answer crawler: one blocks both PerplexityBot and Perplexity-User, the other blocks PerplexityBot only. Blocking these crawlers means disappearing from the answers those tools generate. In our own client work, this is nearly always a side effect of a security or CDN setting rather than a deliberate choice. We did not interview these two companies, and we cannot know their actual intent.

For training crawlers: 2.7% of sites (5 of 182) block GPTBot and ClaudeBot, 2.2% (4 of 182) block Google-Extended, Applebot-Extended, and CCBot, 1.1% (2 of 182) block meta-externalagent, and 3.3% (6 of 182) block Bytespider. Blocking training crawlers is, under our methodology, a legitimate business decision that does not affect whether a brand gets cited in an answer.

Methodological note: this check measures what robots.txt declares, not whether a firewall or CDN blocks those crawlers beneath that layer. We identified our own crawler honestly throughout the run and never impersonated a real bot to test infrastructure-level blocking.

Finding 2: real FAQ structure is rare; loosely detectable FAQ content is more common

Only 1.6% of sites (3 of 184) carry real FAQPage structured data, the format an AI model can extract into distinct, citable question-and-answer pairs with the most confidence. A further 21.7% (40 of 184) have FAQ-labeled content our script could detect only by heading or link text: a human-visible signal, but not one a machine can parse with the same reliability. Combined, 23.4% (43 of 184) show some form of detectable FAQ content. Even so, three out of four sites in this sample still have nothing in either category.

Finding 3: structured data is common, but mostly generic

72.3% of sites (133 of 184) include at least one block of structured data. The five most common types:

  • WebSite (a generic marker for "this is a website"): 119 sites of 184
  • SearchAction (declares that the site has a search box): 93 sites of 184
  • Organization (basic company name/logo data): 93 sites of 184
  • BreadcrumbList (site navigation trail): 69 sites of 184
  • WebPage (a generic marker for "this is a page"): 68 sites of 184

We found 45 distinct schema types across the sample; the full list, with counts, is available on request at hello@sellwithmarketing.com. The most common types above are, for the most part, the ones platforms like WordPress, Wix, or Shopify insert automatically without anyone configuring them on purpose. The types that would actually help an AI model cite specific information about what each company makes (Product, 2 sites; HowTo, 0 sites; FAQPage, 3 sites) are nearly absent from this sample.

Finding 4: one in twelve sites has zero of the signals we measured

Crossing the four main signals (real llms.txt, structured data, FAQ content of either kind, meta description), we found that 8.2% of the sites audited (15 of 184) have none of the four. That means 15 manufacturers, with real production capacity, currently say nothing verifiable to an AI model about who they are or what they do.

An important caveat: we measured the code the server delivers, not what a human sees once the site has fully loaded in a browser. If any of these 15 sites add content with JavaScript after load, a visitor might see FAQ content or information our measurement does not capture. An AI model would not see it either, since most of them do not execute JavaScript.

Finding 5: basic hygiene is fairly strong

81.5% of sites (150 of 184) have a meta description, and 73.4% (135 of 184) have a clear H1 heading on their homepage.

Finding 6: llms.txt, an emerging convention with real, and higher, adoption

35.9% of sites (66 of 184) have a real, verified /llms.txt file. No major AI model has publicly confirmed it consumes this file systematically. The cost of implementing it is close to zero; its value, for now, is optionality, not certainty.

What we conclude

Access is not the main limitation here: even where the door is open, most of these sites have little an AI model can pick up and repeat with confidence. A small but real share of these manufacturers block the very crawlers that generate live AI answers, and real, machine-readable FAQ structure sits at 1.6%, with three in four sites having no detectable FAQ content of any kind. With real manufacturing capacity and, in most cases, no digital structure to translate that capacity into something citable, these companies' competitiveness in an AI-mediated buying process depends today on a variable most of them do not yet know is being measured.

Why this matters if you are evaluating a B2B marketing agency

If you got this far wondering whether your manufacturing company needs a digital growth engine, this is exactly the kind of evidence you should ask any agency for before you hire them: not a promise, a live measurement, with published methodology.

This is exactly the variable we work on every day at Sell with Marketing, and we do not stop at diagnosis: we fix it. We design and build your website with the structure an AI model needs to read it and treat it as a reliable source, and we work on your content so that when a buyer asks ChatGPT, Perplexity, or Google itself who can do what your company does, you have something real for that model to cite. This audit uses Argos, the same methodology we use to measure where a client stands before we build anything: first we measure where your company stands today, then we build the citable structure that visibility requires.

Write to hello@sellwithmarketing.com. It reaches Manuel García, our founder, directly.