The Answer Layer: Why Specification Content Is What Gets Cited
An answer engine does not place you fifth. It either builds its paragraph out of your page or it leaves you out entirely. For an Edmonton supplier that difference usually turns on something unglamorous: whether the capacities and standards your buyers ask about over the phone are written down where a machine can read them.
A millwright looking for a replacement gearbox no longer types three words and scans ten blue links. He asks a full question — whether a ratio will fit an existing footprint, which shaft configurations are interchangeable — and gets a written answer with a few sources named beside it. Occasionally one of them is a distributor in Nisku who published a compatibility chart four years ago and forgot about it.
That last case is the subject here. Being named inside a generated answer is a different outcome from ranking, produced by different material, and estimated by inference rather than counted. Each of those facts has consequences for what you publish and what you may claim in a report.
When the engine writes the answer instead of the list
A classic results page is a shortlist: ten candidates, and the reader chooses. A generated answer removes the shortlist. The engine reads a set of sources, composes a response in its own words, and names a few pages it drew from.
For a consumer question that is cosmetic. For a technical one it is structural: the engine now does part of the specifying work that used to happen on your site. If your page supplied the number it used, you are named. If your page said "contact our team for details", it had nothing to supply.
| Behaviour | Classic results list | Generated answer | What it changes |
|---|---|---|---|
| What the reader receives | Ten ranked candidates | One response, a few sources named | No second page to be on |
| Unit of competition | Page against nine other pages | Passage against every passage | One strong table can carry a small site |
| What wins | Relevance and authority | A clear, checkable statement | Marketing prose is unusable material |
| Effect of a vague page | Ranks poorly, gets some traffic | Unusable, therefore invisible | "Call us for specs" costs more |
The second row is worth sitting with. In a ranked list a small distributor competes as a whole document, and a thin site loses to a large one almost automatically. In a generated answer the comparison happens at passage level, so a short cross-reference table resolving an obsolete part number can be reached for ahead of a manufacturer's product page.
Cited in the paragraph, or listed underneath
These get discussed as though one were a newer version of the other. A position is a coordinate on a results page for a query in a location on a date, observable by running the query. A citation is an editorial decision made inside a model at the moment of generation, varying with how the question was phrased and what was asked before it.
So they behave differently under repetition. Run the same rank check twice an hour apart and expect the same number. Ask an answer engine the same question twice and you may get two different sets of sources, both reasonable. That is not a fault; it is what the system does.
- Ranking is observable, citation is sampled. You can check a position. With an answer engine you can only ask repeatedly and record the result — a sample, with the error that implies.
- A strong rank does not guarantee a citation. Pages that rank first on broad head terms are often the least specific, which makes them poor raw material for an answer.
- A citation does not guarantee traffic. The reader may have what he needed. You gained your name in front of a specifier — real, but not a session in analytics.
- Both still depend on being indexed. Nothing can be cited that was never crawled, so the work underneath does not become optional.
What the machine reaches for is whatever states a number
The pattern is consistent enough to plan around. Answer engines reach for material that states something checkable: a capacity, a tolerance, a standard, a compatibility, a procedure with steps in order. They avoid material asserting a quality with no figure behind it. "Heavy-duty construction" cannot be used in an answer. "Rated to 12,000 kg at 1.5 m load centre" can.
That is good news for Edmonton. The city's suppliers are frequently indifferent at marketing copy and excellent at specification, and what an answer engine wants is what a shop already produces for its own operations — then treats as internal.
The specification table
Capacities, dimensions and operating limits in rows and columns, model number in the first.
- Units in every cell
- One row per configuration
- Standards named
The cross-reference chart
Your number against the one the buyer has: a competitor's, a superseded one, an OEM part on a machine you did not sell.
- Both directions, one page
- Fitment notes where a swap is conditional
- Real text, not an image of a table
The materials guide
Which alloy, grade or coating for which service condition, with the trade-off stated.
- Named grades, not "premium steel"
- Failure modes stated
- Temperature and corrosion limits
The procedure page
How an inspection or commissioning step is carried out, in order, with the governing standard cited.
- Numbered, self-contained steps
- Standard and revision year named
- Who is qualified to sign it off
Each answers a question in a self-contained way. A model retrieving a passage takes a chunk; it does not read your navigation or the paragraph three screens above. If the chunk needs context that is not in it — a unit given only in the page heading, a model number only in the breadcrumb — it is often unusable.
The specification content you have, and where it is buried
Before commissioning anything new, run an inventory. In almost every machine shop, equipment dealer or technical services firm, the material answer engines want already exists — just not in a form anything can read.
| Where the specification lives | Typical example | Why a machine cannot use it | What publishing takes |
|---|---|---|---|
| A scanned PDF catalogue | A manufacturer's 180-page guide, uploaded in 2016 | Images of tables, no text | Retype the pages that matter as HTML |
| A sales engineer's head | Which pump curve suits which slurry | Never written down anywhere | Two recorded hours, edited into pages |
| The quoting spreadsheet | Capacity, reach and material limits | Internal file, never a page | Export the columns a customer may see |
| The reply email template | The standard answer to "can you do this in 316?" | Sent one to one, never indexed | One page per recurring question |
| A supplier's dealer portal | Full fitment and supersession data behind a login | Not crawlable, not citable | Publish what your agreement permits |
The fourth row is the cheapest win most firms have. A fabrication shop answers the same fifteen questions every week: plate thickness, bed size, certification to a particular division, lead time on a stainless job, whether it can weld duplex. Those answers already exist in sent mail, precise enough to quote from. Publishing them costs an afternoon.
The framing that keeps this useful is deliberately narrow. The question is not "how do we get the answer engine to favour us", which nobody outside the model vendors can answer honestly. It is "is what our buyers ask us on the phone written down anywhere a machine can read". That one you can answer today, with a list.
How a visibility estimate is assembled, and what it rests on
Since no counter exists, tools build an estimate, and its construction determines how far you may push the conclusion. The AI Analytics section of the Semalt panel is a workable example: six views producing a competitiveness picture rather than a count of citations.
Generative market research, not a citation log
What the module produces, so you can judge what to do with it.
- A competitiveness score with a Market Circle. Competitors sorted into top-tier, mid-tier and niche bands, showing who you are actually compared against.
- Model-generated market context. Positioning, a traffic estimate and opportunities — how a model characterises your business when asked.
- Query research and intent classification. Questions grouped by what the asker wants to accomplish, which matters more here than volume.
- Pages flagged as leverage. Existing URLs marked for content expansion or internal linking, rather than advice to write from nothing.
- Competitor strengths and content gaps. Where rivals are strong and where a subject is covered by nobody — for specification work, a publishable list.
- A global visibility value across the AI landscape. One aggregate figure for the portfolio. Useful as a trend line, not as a measurement.
Notice what is absent: no line saying "cited 41 times last month". What you get is a model's characterisation of a market and your position in it, expressed as scores and bands. Useful for deciding what to publish next — and a different category of thing from a metric.
Content suggestions, and publishing what you already say
The output of that analysis is a list of things to publish. The tiers matter mainly for who turns it into pages and who checks them first — a bigger question in technical fields, because a wrong figure in a specification table is not a typo.
Automated suggestions, or suggestions with a review gate
For specification content this decides whether the output is usable.
- AutoSEO, 149 USD per month per domain. Automatic keyword discovery and prioritisation, automated link building, on-site AI suggestions, full Search Console and SERP analytics, live AI chat.
- FullSEO, 500 USD per month per domain. Adds manual keyword selection with automatic fallback, manual link placement against a target domain rating, and human review of on-site changes.
- The review mode decides it. Nothing stating a capacity or a certification should go public without someone who knows the shop reading it.
- Every keyword candidate is approved, rejected or deferred. That is how a technical pool stays free of terms describing work you do not do.
Whichever tier you run, judge suggestions by one test: does the page state something specific enough to quote. "Five things to consider when choosing a gearbox" states nothing. A page listing ratios, mountings and service factors for the six gearboxes you stock states six things, each liftable into an answer.
The general advice post
A rewrite of what everyone publishes. Nothing checkable, so nothing quotable.
- No figures
- No standards named
- Indistinguishable from ten rivals
The written-down phone answer
One page per question your counter staff answer weekly, in their words, numbers included.
- Exact capacities and limits
- Conditional cases stated
- Written once, reused in quotes
The gated spec sheet
A form between a specifier and a dimension. It converts nobody early in a cold cycle, and cannot be crawled.
- Invisible to retrieval
- Loses the citation to a rival
- Filters out the researcher
The open table, gated extras
Publish the specification; keep CAD files, drawing sets and configured quotes behind the form.
- Citable public layer
- A real reason to fill the form
- Enquiries arrive better informed
None of this needs new expertise. It needs a decision that the answers your people give twenty times a week are a published asset rather than a private one — the same argument behind our technical SEO for catalogue sites.
Folding this into ordinary reporting without overclaiming
The reporting problem is easy to describe and easy to get wrong. Two categories of number sit side by side: counted events from Search Console and rank tracking, and inferred estimates from generative analysis. One table for both quietly claims they are the same kind of thing.
The fix is separation and labelling. Counted metrics — clicks, impressions, click-through rate, average position, keyword dynamics tracking entries into and exits from the top three, ten and thirty — go in the results section; inferred figures go in their own. The configurable report builder handles the mechanics: CSV and JSON to 10,000 rows, server-rendered PDF to 250 rows, branding with your logo and colours.
- Two sections, never one table. "Measured performance" and "AI visibility estimate", each with its own heading, so nobody guesses which is which.
- Give the estimate a stated method. One line saying it is model-inferred, from sampling, not a citation count. Written once, reused monthly.
- Report direction, not decimals. "Up over two quarters" is defensible. "Up 3.4 % since March" implies precision the method cannot support.
- Tie it to an action, not a result. The score prompts "publish the fitment tables next"; counted metrics report whether that worked.
- Watch the counted proxies. Impressions on question-shaped queries and traffic to spec pages are real numbers that move when this work lands.
Questions that come up first
Does this replace what we are already doing for search?
No, and it depends on it. A page never crawled cannot be retrieved, ranked or cited, so indexing coverage, site speed and a clean catalogue structure remain the foundation. What changes is which content earns the extra outcome on top: checkable, self-contained passages rather than persuasive prose.
Can we find out how often we are actually cited?
Not exactly, and not from any official source. You can sample: ask an answer engine the twenty questions your buyers ask, on a fixed schedule, and record which sources it names. That is defensible as long as it is called sampling. It is not a count, and results move between runs.
Our competitors publish nothing technical. Advantage or warning?
Usually an advantage, provided the questions are being asked. If nobody in your category has published the compatibility data, the first firm to do it becomes the material every answer is built from. Check the content-gap view first: a gap with no demand behind it is an empty room.
Are we giving away information competitors will copy?
Capacities, tolerances and standards are not trade secrets — they are what you tell every caller. Your process, tooling, pricing and drawing sets are the assets, and they stay behind the form. Firms that hide specification sheets are not protecting anything; they are just harder to specify into a project.
How long before any of this shows up?
First measurable movement on the counted metrics typically appears after four to eight weeks, and on a long industrial buying cycle the enquiry may arrive months later. The generative side has no reliable published lag, another reason to treat its numbers as directional.
Where does the assistant fit in?
The Stream assistant inside My SEO is bound to real project data: a router decides per question which data blocks to load, answers stream token by token, and up to twenty messages are kept. It takes keyword and URL lists in batches — the practical way to move a list of publishable questions into a plan.
Where a shop floor should start
The sequence is unromantic. Week one: list the fifteen questions your counter and sales engineers answer most often, and mark which are published anywhere on your site — usually two or three. Week two: publish the missing ones, in the words already used in sent mail, figures intact. Week three: convert one scanned PDF table into real HTML. Week four: let the query research and content-gap views order the next month.
To see the underlying views before committing to a plan, sign in and connect your domain to the dashboard — Search Console and SERP analytics populate once the account is linked, and the generative views run alongside them rather than instead. More on catalogue-scale search work sits in the Edmonton SEO blog.
