Assistants answer from text they can retrieve and parse. For an industrial supplier that turns a familiar content problem into a visibility problem, and it rewards specification over positioning.
A specifying engineer looking for sintered filter elements at a particular micron rating used to type that phrase into a search engine and work down a page of results. Increasingly the same question goes to an assistant, and what comes back is a short list of named suppliers with a sentence of justification each. The buyer may never see a results page at all.
This does not remove search. It adds a layer above it that compresses a page of links into three names. For a supplier the consequence is closer to binary than gradual: either the assistant names you in that conversation or you are absent from it.
How each assistant assembles that shortlist differs by vendor and is not published in any detail. What is observable is the raw material. Assistants answer from text they can retrieve and parse, and the suppliers that get named tend to be the ones whose capabilities are stated somewhere in plain, indexable prose.
A model cannot infer that a fabricator handles 6 mm stainless in 3 m lengths from a photograph of a press brake. It works from statements. If the tolerance, the material range, the approvals and the applications exist as text on a page, they are available to it. If they exist only in a sales engineer's head, in a quotation template or in a spreadsheet behind a login, they are not.
This sounds obvious and is routinely ignored. A great many industrial sites describe what the company believes about itself rather than what it can make. The capability page promises precision engineering to the highest standards and never states the machine envelope, the grades run, or the tightest tolerance actually held in production.
The remedy is unglamorous. Write the specification down: machine envelopes, material grades, the sizes and thicknesses genuinely run rather than the ones theoretically possible, the scope of each approval, and the , named in the terms those sectors use themselves.
Much of the real specification in an industrial business lives in PDFs: data sheets, catalogues, approval certificates, installation manuals. Extraction from those files is inconsistent. Some crawlers pull text out and some do not, quality depends on how the file was produced, and a scanned or heavily laid-out document may yield very little usable text. A parameter table set in print-oriented columns often comes out of extraction scrambled beyond use.
Treat the PDF as the download, not as the record. The same specification should also exist as an HTML page: one product or product family per URL, parameters in a real table with header cells, units stated explicitly, and the PDF linked from it for whoever wants the printable version.
This costs effort, and for a catalogue of several hundred line items it is a data problem rather than a copywriting one. The sensible route is usually to publish pages from the product database instead of hand-building them. Where a client cannot do the whole catalogue, do the range that actually gets quoted.
Schema.org markup gives a machine an unambiguous reading of things a page states ambiguously. Three types carry most of the weight for an industrial supplier, and they are worth implementing properly rather than partially.
The limit matters as much as the benefit. Structured data describes; it does not persuade and it does not confer authority. Marking up a thin page as a Product does not make it a better source, and no markup obliges any system to use it. Treat schema as a way of removing ambiguity from content that already carries information, not as a lever that moves rankings on its own.
Most robots.txt rules and bot-management settings on industrial sites were written to deter scrapers and reduce load. Several assistant vendors now run distinct user agents for different purposes: one that gathers training data, one that fetches pages live while answering a question, and one that feeds a search index. Blocking them remains a legitimate choice, but it is no longer free.
The distinction worth explaining to a client is that the agent fetching a page to answer a buyer's question right now is not the same as the agent collecting training data. A blanket disallow can remove the supplier from live answers while doing little about the training question, which earlier crawls had largely already settled.
In practice, read the robots file and the CDN or WAF bot rules together, because a bot-management product will often be blocking agents the robots file permits, and nobody in the business will know. Decide per agent, write the reasoning down, and review it when a client's competitors start appearing in answers and they do not.
Before advising anything, get a baseline. The exercise is cheap and takes an afternoon.
The caveats are serious. Outputs vary between runs on identical prompts, they are influenced by account history, and they change without notice when a model is updated. This is reconnaissance, not measurement, and it should not be sold to a client as a metric. What it does tell you reliably is which sources an assistant reaches for in a category, and that is often a trade directory, a standards body or a competitor's technical page. The list of cited sources is more actionable than the ranking.
Buyers ask in the vocabulary of the specification: alloy designations, standard numbers, tolerance classes, connection types, duty and ingress ratings. A system matching a question to text has something to match against when the page uses that vocabulary. Marketing abstraction has no counterpart in the question and contributes nothing to the match.
So write AISI 316L, ANSI 150 flange, IP66, Class 8.8, rather than premium-grade materials and rugged construction. Include the synonyms a buyer might use, the regional variants of a term, and the superseded designation alongside the current one wherever both are still in circulation on the shop floor.
There is a useful side effect. Producing this text forces tacit sales knowledge into writing, and the person best placed to draft it is usually an applications engineer rather than a copywriter. The engineer's draft, lightly edited, is nearly always better raw material than a polished page written from a brief.
Nobody outside the vendors knows in any detail how these systems select sources, how heavily each signal weighs, or how stable that weighting is between model versions. Anyone selling a guaranteed method for AI visibility is claiming knowledge that is not publicly available. Say so to clients, plainly, before someone else sells them the certainty.
What is defensible is the direction of travel. An assistant needs parseable, specific, retrievable statements of capability in order to name a supplier at all. That happens to be the same thing a specifying engineer needs when comparing three quotations, which is why the work is worth doing regardless of how the retrieval mechanics settle.
Advise clients to spend on the durable half of the problem: accurate, structured technical content, reachable by crawlers, maintained as the product range changes. Be sparing with anything that only pays off if one vendor's current behaviour holds, because that behaviour has already changed several times and will change again.
Key takeaways
There is no good evidence that they replace it outright. What is visible is a layer forming above search, where the assistant compresses a page of results into a handful of named suppliers. Buyers still verify on the supplier's own site, which means the site has to answer the technical question when they arrive.
It depends which agent. The crawler that fetches a page live to answer a buyer's question is not the same as the one gathering training data, and blocking both with one rule removes the supplier from answers without addressing the training concern. Decide per user agent, and check the CDN or firewall rules as well, since those often block more than the robots file does.
It makes a page unambiguous to a machine, which helps when the page already states real specifications. It is not a ranking lever and no system is obliged to use it. Marking up a page that contains no technical detail achieves nothing.
Ask ten to fifteen buyer-phrased questions in two or three assistants from a logged-out session, and record both the suppliers named and the sources cited. Repeat it weeks later, because answers vary between runs and change when models are updated. Use it to understand which sources the assistant trusts in that category, not as a score to report.
