Insights
Short pieces on selling engineering, written by our consultants. There are five so far. Two are on how AI tools shortlist suppliers, one on the email fines that rose in February 2026, one on the announcements that show a plant is about to spend, and one on letting AI read a tender pack without touching the price.
What gets a manufacturer recommended by an AI tool
Most business buyers now ask an AI tool at some point before they buy, and the tools cite the pages that state a specification.
In Forrester's 2025 Buyers' Journey Survey, 94% of business buyers said they used AI somewhere in their buying process. In a 2026 Semrush survey of B2B professionals, 41% said they start supplier research in an AI tool before they touch a search engine, and 92% of those who use AI at work said it shaped their shortlist.
The specification, on the page
Gorilla76's audits of manufacturer websites keep finding the machine list, the tolerances and the materials, but in PDF brochures rather than on the page. One firm in those audits had 39,000 words in PDFs and 8,000 in HTML. AI tools read web pages and rarely open brochures, so a table of machines and capacities on an ordinary page does more than a well-designed PDF.
The direct answer comes first
SparkToro's 2026 analysis of where citations fall put 44% of them in the first 30% of a page's content, so the opening carries more than its share of the weight. A page that opens with "a passion for quality" gives the tool nothing. One that opens with "We hold ±0.01 mm on turned parts up to 366 mm diameter, in batches of 5 to 500" gives it a sentence it can quote.
Headings in the buyer's words
"What tolerances do you hold?" and "Which materials do you machine?" match the question the tool has just been given; a heading such as "Precision engineering" matches nothing.
Structured data
Structured data is a set of labels in the page's code that tell a machine what each item is. Organization markup tells it who the firm is, Service markup what it sells and FAQ markup what it has already answered, and none of it changes what a reader sees. In the same Gorilla76 audits, one site carried 1,700 structured-data statements and none of those three.
The llms.txt file is unproven
The llms.txt file is a plain-text summary of a site that some agencies now sell as a way to brief AI tools. It has been promoted heavily during 2026, and we have not found evidence that it changes whether a page is cited, so we do not offer it.
The commercial case is the conversion rate. Seer Interactive's 2025 analysis of AI referrals found that visitors arriving from an AI recommendation convert at a higher rate than visitors from organic search. Our reading of why is that they arrive already shortlisted. The first fix is usually the capability pages, the services or what-we-do pages on your site; the Capability Page Checklist lists what each should state.
Sources: Forrester, Buyers' Journey Survey, 2025. Semrush, B2B AI research survey, 2026. Gorilla76, AI search audits of manufacturers, 2026. SparkToro, citation position analysis, 2026. Seer Interactive, AI referral conversion, 2025.
Cold email to other businesses after February 2026
The maximum fine for breaking the email rules rose in February 2026 and the rules themselves stayed as they were. This piece sets out what a manufacturer may still send to other businesses, and how we send it.
On 5 February 2026 the maximum fine for breaking the UK's email marketing rules rose from £500,000 to £17.5m or 4% of worldwide annual turnover, whichever is higher, under the Data (Use and Access) Act 2025. The rules on what you may send are the same as before; the penalty for breaking them is much larger, and the Information Commissioner's Office (ICO) has said further direct marketing guidance is on its way.
For a manufacturer writing to other businesses, the position is fairly simple. The Privacy and Electronic Communications Regulations (PECR) treat corporate subscribers differently from individuals. You may email a named person at a limited company, an LLP or a public body about matters relevant to their role without prior consent, provided you say who you are and give them a way to stop you. Sole traders and unincorporated partnerships count as individuals, and personal addresses are out of bounds.
UK GDPR sits alongside PECR, and for business-to-business outreach the lawful basis is normally legitimate interest. The ICO expects you to have thought about it and written it down. The document is called a legitimate interest assessment, and it records what you are doing and why, whether you need to do it that way, and whether the person's interests outweigh yours. It takes an afternoon the first time and should be revisited when the audience or the volume changes.
Separately from the law, the mailbox providers have their own rules, and they now reject mail that breaks them rather than file it as spam. Google and Microsoft require SPF, DKIM and DMARC on the sending domain, a one-click unsubscribe header and a complaint rate under 0.3%. A sending subdomain, warmed up and authenticated, keeps a firm's main domain out of it.
Between them, the law and the providers leave room for a short, specific email to a named person about a project they are running, and nothing bought, guessed, personal or high-volume.
We have built these rules into the way we send, so nobody has to remember them. For each client we write and date a legitimate interest assessment, write only to addresses printed on the recipient's own site or supplied by the client, keep a suppression list permanently, put a one-click unsubscribe on every message, cap the volume at forty companies a month from one domain, and pause sending automatically if complaints pass the providers' threshold. Nothing is sent until you have approved it. This is how Outbound sends. The rules, the sending set-up and twelve templates are written up in the Cold Email Playbook, £95 in the library; the first two sections are free to read.
Sources: Data (Use and Access) Act 2025, commencement 5 February 2026 (Blake Morgan and Clifford Chance briefings). ICO, direct marketing guidance. Google, Email sender guidelines; Microsoft, Outlook requirements for high-volume senders, 2025.
What an AI says when a buyer asks for a supplier like you
Ask an assistant for a supplier like you and read the shortlist it gives, because a buyer who asks the same question sees that list before they see you.
Ask ChatGPT, Gemini, Claude or Perplexity for "UK subcontract machinists who can hold ±0.01 mm in Inconel, ISO 9001, batches of a few hundred" and you get a shortlist in seconds. It is worth doing this for your own company, because the shortlist is built from what is written about you, and mostly from your own website.
If a capability page says "quality precision engineering solutions" and nothing else, the tool has nothing it can quote, so it quotes a competitor whose page lists machines, materials, tolerances and approvals in plain terms.
The fix is ordinary work, and most of it is already in your approved-supplier form. Put the machine list on the site with capacities and axes, the tolerances you hold in production, and the materials, batch sizes, lead times, approvals and sectors, each under a heading a machine can find. This is the same information a procurement manager wants for an approved supplier form, which is why it works for both readers.
The first thing the Sales Engine Audit, from £750, produces is a transcript of what the assistants say about your company today, with the pages to fix in priority order and a rough cost and effort against each. The free Findability Check puts six queries of the same kind to the tools and sends the report, with the transcripts, within the hour; the audit runs the full list and adds the fixes.
The announcement comes before the order
The local paper reports the council approving a plant's extension, the trade press reports the new line being ordered, a job board carries the advert for a maintenance planner, and Companies House records the charge when the money is borrowed. All of it is public, spread across thirty or so sources, and all of it appears before the plant places its orders.
The usual approach to outbound is to build a list by job title and sector and write to it. Reply rates are low, because most of the people on the list are not buying anything this quarter. A plant that has just committed capital has a project engineer with a list of suppliers to contact, a deadline and a budget. Writing only to plants that have announced something in the last ninety days means the email can be about the project that engineer is running rather than a general question about whether they need a supplier.
The practical difficulty is the reading. Finding the few dozen announcements a month that matter to a gearbox manufacturer means getting through several thousand pages of press releases, planning notices and job adverts. That reading is the part we hand to AI. The tool goes through them faster than any person can and flags the handful worth a closer look; a researcher then decides which items matter and writes the email.
We build each client's list from the plants that have announced spending in the last ninety days, find the engineer running the project, and write about that project. That is Outbound on the products page; the Signal Feed sends the announcements and leaves the writing to you.
Using AI in the quoting office without losing control of the numbers
The tools may read the pack and draft the questionnaire answers, and the price and the drawings stay with the estimators.
On a tender pack the reading takes longer than the pricing. A typical pack runs to a couple of hundred pages of specification, drawings, schedules and preliminaries, plus a questionnaire that asks the same thing four different ways. Reading is what current AI tools are good at, and the way to use them safely is to be strict about what they are allowed to touch.
What the tool may do is agreed before it reads anything. It may read the pack and produce a structured summary of the scope, the quantities, the materials, the standards referenced and the deadlines, with every question in the questionnaire pulled into a list, and it may draft answers to those questions from a library of your previous answers. It is not allowed to produce a price or to read a drawing, and every quantity it extracts is checked against the source before it goes anywhere near a spreadsheet.
The library is where the drafts come from, so it gets built first. A day spent pulling the best answers from your last forty tenders into one document, tagged by topic, does more for the drafts than any amount of prompt-writing, because the tool can only draft from what the library holds.
The library has one owner, the senior estimator rather than the managing director or IT, and it is the owner who updates it every time a tender goes out and tells the rest of the team when something changes. Without an owner the library stops being updated and the drafts get worse with each tender.
Set up this way, the estimators' first pass on a pack starts from the summary and the drafted answers rather than from page one, and the pricing and the drawings stay with them. The rule set, the prompts and the rollout plan are written up in the AI in the Quoting Office playbook.
The free checkcomes first
Two of the pieces above come down to what the assistants say when a buyer asks for a supplier like you, and the report shows you that within the hour.
If you would rather talk first, open the booking page.