A fictional UK manufacturer of industrial measurement equipment
- Estate
- About 1,400 published pages on one English-language site
- Sample
- 100 items (10% of the estate is 140, capped at 100)
- Language
- English only
A professional opinion you can act on or argue with. Every finding traces back to the item and the check that produced it.
The question and the decision
Calder & Finch plans to put an AI assistant on its support site next quarter. Before funding that, it wants to know whether its existing content can be found, trusted and reused by AI systems, and what to fix first.
What we reviewed, and what we did not
Sampled items
- 40 product and specification pages
- 30 support and how-to articles
- 20 blog and insight posts
- 10 company, contact and policy pages
Assumptions
- The public site is the content the assistant will draw on.
- Items were sampled across sections in proportion to their size, not chosen for problems.
Limitations
- A sample, not the whole estate: counts describe the 100 items reviewed, not all 1,400.
- Structural and coverage findings only. We did not check technical claims against your products; we flag where your own pages contradict each other.
- English only. Translated versions were not reviewed.
- Not legal or regulatory advice. Anything with a regulatory angle is flagged for your own counsel.
At a glance
- items reviewed
- with at least one finding
- priority findings
- shown failing in an AI answer
In one sentence: the content an assistant would rely on most, support and product detail, is the content least able to show it is current, so fix ownership and dating before launch, then make specifications readable as text.
Top five findings
- Effort: Medium · Owner: Product marketing, with a CMS template change
Key specifications live only inside PDF datasheets
- What we found
- On 26 of the 40 product pages, operating range, accuracy and approval ratings appear only in a downloadable PDF, not in the page text.
- Example from the sample
- The product page for the (fictional) TX-400 pressure sensor describes the product in two paragraphs and links to a datasheet. The operating temperature range is on page 3 of the PDF.
- Why it matters
- AI systems retrieve passages of text. When the answer sits inside a PDF, it is found less reliably, and the answer engine falls back on older or third-party sources.
- What we recommend
- Publish the six most-asked specification fields as structured text on each product page, starting with the 15 highest-traffic products. Keep the PDF as the downloadable copy.
- Effort: Small · Owner: Support team lead
Support articles carry no review date or owner
- What we found
- 28 of the 30 support articles show no last-reviewed date and no named owner. Four give a menu path that differs from the one in your current release notes.
- Example from the sample
- “Calibrating the TX series” still refers to a “Setup > Calibrate” menu. The release notes for firmware 3.2 rename it “Service > Calibration”.
- Why it matters
- An answer engine has no way to tell current guidance from retired guidance. Neither does a customer. An assistant built on these articles would repeat the old instructions with confidence.
- What we recommend
- Add a last-reviewed date and an owning team to every support article. Review the four affected articles first. Set a review interval for articles that mention firmware.
- Effort: Small · Owner: Web team (template change)
Headings are used for styling, and answers are hidden in tabs
- What we found
- 41 items skip heading levels or use headings as visual labels. On 19 product pages, the answers to common questions sit in collapsed tabs with generic labels such as “More”.
- Example from the sample
- A product page moves from its title straight to a fourth-level heading reading “Why choose us”. Installation guidance sits under a tab labelled “Details”.
- Why it matters
- Headings tell people and machines what a passage is about. When they describe styling rather than content, retrieval systems split and label the page badly, and the right passage is not returned.
- What we recommend
- Fix the heading order in the product-page template, not page by page. Rename tabs after the question they answer, for example “Installation” or “Compatibility”.
- Effort: Medium · Owner: Product marketing, reviewed by applications engineering
No page helps a buyer choose between product series
- What we found
- None of the 100 items compares the TX, TR and TM series, although all three are sold for overlapping applications.
- Example from the sample
- A buyer asking which series suits a high-humidity environment has to open three product pages and two datasheets and compare them unaided.
- Why it matters
- Comparison questions are where buyers go to AI tools first. With nothing on your site to draw on, the answer comes from a distributor or a competitor.
- What we recommend
- Publish one selection guide covering the three series against the five criteria buyers ask about most. Link it from each product page.
- Effort: Small · Owner: Support and marketing jointly, one decision needed
The same guidance is published twice and has drifted apart
- What we found
- 7 support topics also appear as blog posts. In 5 of those pairs the steps differ, and neither version says which one is authoritative.
- Example from the sample
- The support article on cleaning the TR sensor head recommends isopropyl alcohol. The 2023 blog post on the same task recommends a mild detergent solution only.
- Why it matters
- Every duplicate doubles the maintenance work and gives an AI system two conflicting sources. The problem grows each time a new team publishes on a topic that already has an owner.
- What we recommend
- Make the support article the single source for any procedure. Point the blog posts to it, or retire them. Agree which team owns procedural content.
Shown failing in an AI answer
We put buyer and customer questions to a general-purpose AI answer engine and recorded what came back. Two examples, each traced to a finding above.
The answer gave a range from a 2019 distributor listing, which differs from the current datasheet. It did not cite the Calder & Finch site.
The current range exists only inside the PDF datasheet (finding 1), so the distributor page was the most retrievable text source.
The answer quoted the “Setup > Calibrate” steps from the support article, a menu path that no longer exists on current firmware.
The support article is undated and unowned (finding 2), and nothing on the page signals that newer guidance exists.
AI answers vary between runs and over time. We record the date, the question and the answer so the result can be checked again after changes.
What to fix first
Findings 2 and 5: date, own and de-duplicate support content
Small effort, and it removes the most likely wrong answers before the assistant launches.
Finding 3: fix the product-page template
One template change improves every product page at once.
Findings 1 and 4: specifications as text, plus a selection guide
More effort, but this is where buyers' questions are going unanswered today.
What this Snapshot did not cover
- The other 1,300 pages, translated content, and content behind a login.
- Whether technical claims are accurate against your products.
- Choice of AI vendor or model, or how the assistant should be built.
- Rewriting content. We recommend what to change; your teams or a separately agreed scope make the change.
Readout, handover and your next decision
The report comes with a 45-minute recorded readout with the person who wrote it. After that, the decision is yours.
Use the findings as they stand. The report and the readout recording are yours to keep.
The priorities above are written so your own teams can run them.
A Full Estate Audit covers the whole estate, with a costed remediation roadmap. 100% credited toward a Full Estate Audit if you sign within 30 days.
Illustrative sample: a fictional company and synthetic findings. A real Snapshot reviews your own content and reports only what that review found.