The question, not the keyword
A buyer now types a whole decision into a chat box: constraints, budget, certification. The unit of work is the answer to that sentence, not a page optimised for one phrase.
Skip to content
Industry solutions
One method, four different question sets
The method does not change between industries. What changes is the questions buyers ask, the facts that answer them, and where those facts have to be published before a model will use them. Below are the four we work in most, with the check you can run on each.
Overseas reference price starts at US$1,800–4,000 per month, minimum term 3 months. Scope and fees are set per market, language and content volume in a written proposal.
Your own page has to hold the fact
An assistant paraphrases what it can read. If the specification, the certificate number or the lead time is only in a PDF attachment, it is effectively unpublished.
Placement is corroboration
One page saying something is a claim. A page plus a directory plus a trade body record saying the same thing is a fact the model is willing to repeat.
Four industries, four starting points
Each section lists the questions buyers in that industry actually ask, what we build to answer them, and a check you can run yourself. Nothing below is a client result — these are the question patterns and deliverables, not measured outcomes.
E-commerce and DTC brands
Product facts that survive being paraphrased
What buyers actually ask
- Which brands sell a [product] that works with a 220V supply and ships to the EU?
- What is the actual difference between model A and model B?
- Is this brand's warranty handled locally, or does the buyer ship back to the factory?
Wording varies by market and by assistant. The point of writing the question down is that it becomes testable: you ask the same thing every round and compare the answers.
What we build
- 1
Product and Offer data matched to the spec sheet
JSON-LD for each SKU with the same values as your specification table — dimensions, materials, voltage, certifications, price range. Where the two disagree, the page is wrong, not the model.
- 2
Comparison and specification pages
Pages written for a buyer who has not chosen yet: how two models differ, what the spec means in practice, which variant fits which use. These are the pages a model can quote instead of guessing.
- 3
Consistent brand facts off-site
Marketplace listings, review platforms and distributor pages carrying the same claims and the same model names as your own site. Divergence here is what produces confidently wrong answers.
How you check it
Ask two different assistants the same specification question. The values should match your spec sheet; if one of them invents a number, the fact was not stated clearly enough on the page.
Smallest useful first round
one market, one language, a 12-question set, four pieces of content a month. Anything narrower than this produces one round of data with nothing to compare it against.
Manufacturers and B2B suppliers
Certifications and capacity that can be verified, not assumed
What buyers actually ask
- Who can supply [component] with ISO 13485 and a minimum order under 5,000 units?
- What is this factory's lead time, and which export markets has it shipped to?
- Which manufacturers offer CE-marked models with a documented test report?
Wording varies by market and by assistant. The point of writing the question down is that it becomes testable: you ask the same thing every round and compare the answers.
What we build
- 1
Organisation data with the facts buyers screen on
Organization and LocalBusiness data covering certification numbers, plant capacity, MOQ, lead time and export terms. A purchasing team screens on these five fields before it reads anything else.
- 2
Process and capability pages
What the plant actually does: process steps, quality control regime, tooling, packaging, Incoterms. Written so a buyer's technical reviewer can check it, not so a brochure reads well.
- 3
Profiles on the directories buyers already use
Industry directories, trade association member lists and certification-body registries, kept consistent with your own pages. Third-party presence is what a model treats as corroboration.
How you check it
Open the page source and compare each certification number against the certificate PDF. If the number on the site cannot be matched to a document, treat it as unstated.
Smallest useful first round
one market, one language, a 12-question set, certification and capability pages first. Anything narrower than this produces one round of data with nothing to compare it against.
Education providers
Course facts a prospective student can act on
What buyers actually ask
- Which online Mandarin programmes accept complete beginners and issue a certificate?
- How many taught hours per week is this programme, and what time zone are the classes in?
- Which providers are accredited, and by whom?
Wording varies by market and by assistant. The point of writing the question down is that it becomes testable: you ask the same thing every round and compare the answers.
What we build
- 1
Course data with the fields students filter on
Course structured data with teaching mode, language of instruction, duration, entry level, schedule and the credential awarded. These are the fields an assistant needs before it can recommend anything.
- 2
Programme pages with entry requirements and outcomes
Who the programme is for, what it assumes, what a graduate can do afterwards, what it costs. Vague programme pages produce vague answers — usually the competitor's.
- 3
Third-party references from accrediting bodies
Accreditation records, examination board listings and education media coverage, each pointing back to the programme page it describes.
How you check it
Ask an assistant about tuition, duration and the credential. Then open the programme page and see whether a reader could have answered from it. If not, the gap is on the page.
Smallest useful first round
one programme family, one language, a 12-question set, course data first. Anything narrower than this produces one round of data with nothing to compare it against.
Medical device makers
Regulatory facts stated exactly, with the source attached
What buyers actually ask
- Which suppliers hold a 510(k) clearance for this device class?
- What is the intended use statement, verbatim?
- Which of these devices are CE marked under MDR, and for which class?
Wording varies by market and by assistant. The point of writing the question down is that it becomes testable: you ask the same thing every round and compare the answers.
What we build
- 1
Device-level data with certification identifiers
Structured data carrying clearance or registration numbers, device class, intended use and the standards a device is tested against. Generic marketing copy is worse than nothing here — it invites a model to fill the gap.
- 2
Compliance-reviewed technical content
Technical descriptions, indications and compatibility notes reviewed against the cleared labelling. Claims stay inside the cleared indication; anything outside it does not get published.
- 3
References on regulatory and clinical sources
Regulator databases, notified body records and clinical literature, so a claim can be traced to a document rather than to your own homepage.
How you check it
Take each claim on the site and trace it to a certificate, a clearance entry or a test report. A claim you cannot trace is a liability, not an asset.
Smallest useful first round
one device family, one market, a 12-question set, regulatory data first. Anything narrower than this produces one round of data with nothing to compare it against.
Where the four differ
The comparison below is about scope, not about price tiers. The first thing to build is different in each industry, and that is what usually decides whether the first round produces anything usable.
| Industry | Buyer's question | Core deliverable | Build this first |
|---|---|---|---|
| E-commerce and DTC | Which model fits my setup? | Product and Offer data, comparison pages | Specification table cleaned up, then 4 pieces of content |
| Manufacturers and B2B | Can this supplier actually do it? | Organisation data, process pages, directory profiles | Certification and capability pages |
| Education | Is this programme right for me? | Course data, programme pages, accreditation references | Course data across the programme family |
| Medical devices | Is this device cleared for my use? | Device data, compliance-reviewed content, regulator references | Regulatory data and intended-use statements |
Deliverable lists are scope descriptions, not fixed packages. The question count, platform list, languages and review frequency are agreed in writing before work starts.
How this differs from the SEO you already buy
Same discipline, different surface. If you already run an SEO programme, the useful question is not which one to keep — it is which part of the work is shared and which part is genuinely new.
| Aspect | Conventional SEO | GEO |
|---|---|---|
| Where the result appears | Links on a results page, ranked | Sentences inside an answer, attributed to a source or not at all |
| What is being optimised | Whether a page matches a query | Whether a specific fact can be retrieved, stated correctly and traced |
| What decides inclusion | Crawl, index and ranking signals | Clarity of the fact on your page plus corroboration elsewhere |
| What you can verify | Impressions, clicks, average position | The saved answers to a fixed question set, round over round |
| Time to first signal | Days to weeks for indexing, longer for ranking movement | Depends on crawling and on when a model refreshes — not on a supplier's schedule |
| What nobody can guarantee | A position | A citation, a ranking inside an answer, or a date |
The two share the groundwork: a crawlable site, clean page structure and factual accuracy. Where they separate is the unit of measurement — a page position versus a recorded answer.
What we will not claim
Most of the disappointment in this category comes from promises made at the sales stage. These are the five we decline to make, in any industry.
- That a specific assistant will start citing you, or by when. Model behaviour changes without notice, and no supplier controls it.
- That a placement will be accepted. Directories and publications apply their own editorial rules, and some will decline.
- Positions or rankings inside an AI answer. There is no equivalent of a search ranking you can be moved up.
- Clinical, financial or performance claims beyond what your own documentation supports. We will not write a claim you cannot defend.
- A fixed number of questions covered for a fixed fee. Question count is the largest single cost driver, so it is written into the proposal and the contract.
What we do commit to is the record: the same question set asked the same way each round, the original answers saved, and a log of what changed between rounds. You can audit that record without our help — see the deliverables page.
Not sure which of the four you are?
Send one sentence describing what you sell and which markets you sell into. If your case does not fit any of the four, we will say so rather than reshape it to fit a package.
We answer in English. Invoicing currency and payment terms are specified in the proposal.