A shipper who used to phone three carriers now asks an assistant, and gets back a paragraph naming two of them. If your company is not in that paragraph, nothing on a rank report explains why — and no counter anywhere tells you how often it happens.
Owner-run freight and food companies in Manitoba have mostly noticed this second-hand. A logistics coordinator in Regina, a purchasing manager at a grocery chain, a broker looking for reefer capacity out of Winnipeg — they still search. What comes back is increasingly a written answer naming two or three companies, rather than ten links they work through themselves.
The ranked list has not disappeared. It sits below the answer, second on the page rather than first. That creates a layer above classic ranking: not where your page sits, but whether the machine used it. The AI Analytics section inside the rebuilt Semalt panel surveys that layer with the instruments that exist, and this article is as much about what they cannot do as what they can.
The question got longer, and the answer got written for you
Compare two versions of the same enquiry. The old one was three words: reefer carrier Winnipeg. The new one is a spoken sentence. Who runs temperature-controlled loads from Winnipeg into Saskatchewan weekly, has CFIA-compliant handling, and can take a partial pallet order on short notice.
That sentence is not a keyword. It carries a lane, an equipment type, a compliance condition and a service constraint, and no page is titled after it. The model does what a dispatcher would: it reads several sources, pulls the part that answers each piece, and writes something coherent. Your company either appears in that assembly or it does not.
- Demand did not fall, the click did. The shipper still needs a carrier. They do less looking, because two sentences on screen already narrowed the field.
- The remaining visits are heavier. Someone who clicks after reading a generated shortlist is checking you out, not browsing. Fewer sessions, each one closer to a phone call.
- Position holds while attention drops. Sitting fourth is worth less when the top of the page answers the question and never mentions you.
This is why a flat click-through rate with rising impressions is such a common and confusing reading on a carrier's Search Console. Nothing broke. A storey was added above and your floor stayed where it was.
Ranked and cited are not the same achievement
The two move together to a degree — a page nobody can find is a page no model reads. But the mechanics differ enough that one number for both will mislead you.
A ranking is a queue: one first place per query, and what you take somebody else loses. A citation is not. One answer draws on four sources at once, and admission depends on whether a passage closes part of the question checkably. A page sitting eighth gets used because one paragraph states transit times to Thunder Bay in days. A page sitting second gets skipped because it says "reliable service you can count on", which a machine can do nothing with.
| Question | Classic ranking | The answer layer |
|---|---|---|
| What is counted | A position against a keyword | Whether a passage was used to build an answer |
| How many can win | One occupant per slot | Several sources sit in the same answer |
| What earns the place | Relevance, links, a crawlable site | A checkable statement, tightly scoped |
| Where the number comes from | Search Console and rank tracking, daily | Inference from proxies, no official source |
| Can you attribute a change | Usually, imperfectly | Rarely, because model updates are undated |
| Who your rival is | Other carriers and platforms | Whoever wrote the clearest sentence on the topic |
The bottom three rows are the uncomfortable ones. Granularity differs too: a rank belongs to a keyword, a citation is earned by a claim. One paragraph on frozen-goods handling can turn up in answers to twenty phrasings without the page ever leading a list.
The answer is built out of platforms you already lose to
Here is the part that lands badly in a Manitoba trucking office. When a model assembles an answer about who hauls a lane or who runs cold storage in the province, it does not primarily read carrier websites. It reads whatever describes carriers at scale: load boards, freight directories, association member lists, review sites, business registries, trade coverage. Those are the same platforms that already outrank you for your own service terms.
Being cited mostly means being described somewhere a machine already trusts. That is a different job from ranking your own site, and it is not entirely in your hands.
Directories, load boards and listing platforms
The layer that carries most of the raw description of who does what in freight.
- They are structured, and that is why they win. A directory entry states equipment types, service area and authority numbers in fields. A model lifts that without interpretation.
- Your listing is often years old. The lanes and certifications on those profiles were entered by someone who left in 2019. Correcting them is the highest-yield hour available.
- Consistency beats polish. Three profiles saying three different things about your service area give the model nothing firm to state.
The pages you actually own
Smaller in volume, but the only place you can write the sentence yourself.
- Specifics beat description. "Dry van, reefer and flatbed, 42 power units, weekly service to Saskatoon and Thunder Bay" is quotable. "Full-service transportation solutions" is not.
- Your site is the tiebreaker. When platform data disagrees, a current page on your own domain is what settles it.
- Volume is not the lever. Six accurate pages outperform sixty vague ones, which suits a company with no marketing staff.
How a visibility estimate is built, step by step
Since nobody can count citations directly, the tools model them, and the construction is what limits the conclusion. In the panel this sits in six AI Analytics views, built alongside Search Console and SERP data for the same domains.
The first view scores a domain's competitiveness and draws the field around it as a market circle, sorted into top tier, mid tier and niche. For a Winnipeg carrier the top band is national fleets and the platforms themselves. That band is a map of how the category is organised, not a target. The mid band is where companies your size sit. The niche band teaches tactics: a nine-page site that explains one thing thoroughly turns up in answers month after month.
The second view has a model describe your domain back to you — positioning, a rough traffic estimate, openings it sees. Its value is the outside reading, not the numbers. A company that considers itself a temperature-controlled specialist often finds four fifths of its pages describe general freight, and the description says so. That contradiction is the output worth having.
The rest handle query research with an intent classification, competitor strengths and coverage gaps, pages flagged as leverage for expansion or internal linking, and one compressed visibility figure across the portfolio. The AI research views in the panel sit next to the SERP competitor list, so you can put the two sets of rival names side by side. Where they disagree, the disagreement is the finding.
What the estimate is good for, and where it must not go
A rough instrument is not a useless one. A fuel gauge reading a quarter low still tells you when to stop. The discipline is knowing which decisions it can carry.
As a direction over quarters
Three or four readings across a year say something about direction. One reading alone says almost nothing.
- Compare only with your own history
- Expect noise between readings
As an alarm to investigate
A sharp fall is a reason to open the underlying views and look at coverage. It is not a conclusion and never a headline.
As a number in a customer document
It has no business in a rate quote, an RFP response or a shipper's vendor scorecard. Those are read as commitments.
- Use audited figures there
- Keep estimates internal
As a target in a contract
Nobody controls the model updates that move it. In an agreement, it makes an outside party's release schedule your obligation.
What makes a passage usable as a source
There is no formula, but the same properties keep showing up in pages that get used, and most of them are what good trade writing always looked like. What changed is the penalty for vagueness: once a stylistic weakness, now a missed mention.
Numbers with units
Trailer counts, pallet positions, dock doors, transit days, temperature ranges. A machine can restate a number, not an adjective.
One question per heading
Put the question in the heading and the answer directly beneath it. An answer spread over four paragraphs rarely reads as a finished claim.
Boundaries stated
Which lanes you do not run, which commodities you refuse, minimum order sizes. Exclusions are unusually citable because almost nobody publishes them.
A visible date
Rates, capacity and certifications change. A page showing when it was last reviewed is easier to rely on than one that could be from any year.
The content suggestions in the panel propose expansions and internal links against pages you already have. For an owner-run company the instruction is to ignore anything implying a publishing schedule and take only the rewrites: three thorough revisions of existing service pages beat twelve thin new ones, and they fit into a week of evenings.
| Gap | How it shows up | First move | Effort |
|---|---|---|---|
| No stated service area | Answers name carriers who list provinces explicitly | Write the lanes and provinces out, including the ones you decline | low |
| No equipment detail | Competitors list trailer types and counts, you say "modern fleet" | Publish the fleet mix with numbers and a review date | low |
| No certification specifics | Cold-chain and food-grade answers cite whoever names their standards | Name each certification, its scope and its renewal cycle | medium |
| No process description | Niche sites explain cross-dock or claims handling end to end | Document the actual sequence on the service page | medium |
Why this lands badly in an owner-run company
For a family carrier with forty trucks, or a food processor whose website was built once and left alone, this is not a growth hack. It is closer to a diagnosis. The machine describes you using other people's descriptions, most of them sit on platforms that compete with you for the same enquiries, and several are out of date because nobody was assigned to maintain them.
Two consequences follow, neither comfortable. The fastest available improvement is administrative rather than creative: fixing listings, standardising how the service area is stated, matching certification names across profiles. And some of this cannot be fixed by you at all. If the model prefers a load board's page to yours, you will not outrank the load board on its own ground, and pretending otherwise wastes money.
One structural advantage sits on your side. A carrier selling out of Winnipeg into Saskatchewan, Ontario and the northern states has a genuinely specific service story, and specificity is what this layer rewards. The national fleet's page has to be vague because it serves everywhere. Yours does not.
Folding this into the reporting you already do
You do not need a separate AI report. You need one section, labelled as an estimate, inside the report you already read monthly or quarterly. The rest stays as it is: clicks, impressions, average position, and which keywords entered or left the top three, top ten and top thirty.
Put the measured material first and the inferred material last, with a line between them saying which is which. Export the same tables as always — CSV or JSON up to ten thousand rows, PDF up to two hundred and fifty — and add one page at the back.
- Measured, first section. Clicks and impressions, position changes, top-30 entries and exits. The part that can be checked, and the only part that belongs in a customer conversation.
- Inferred, last section, labelled. The visibility figure with its previous two readings beside it, and a one-line note that it is a model estimate with no official source.
- Actions, not numbers. What was fixed, what is next, what was deliberately left alone. An owner reads that page and skips the rest, which is fine.
Cadence matters more than depth here. Quarterly is honest for this layer, because a monthly reading of an estimate mostly measures noise. Keep the classic figures monthly — the Search Console and SERP views refresh on their own — and review the inferred number four times a year, against its own history, in the meeting where you decide which page gets rewritten next. The wider approach to measurement is set out across our blog.
Questions we get asked
Is classic ranking finished, then?
No. Transactional and local searches still resolve through the ranked list, and someone looking for a warehouse near their plant still clicks a map result. The shift concentrates on the research questions that come before a shortlist. Dropping position work to chase citations loses on both layers.
Can we pay to be cited in a generated answer?
Not directly, and be careful with anyone who says otherwise. You can make yourself easier to describe accurately: correct listings on the platforms models already read, specifics on your own pages, mentions in places with editorial standing. That is the whole mechanism.
We serve four provinces from one address. Does that hurt us here?
It hurts if your site only talks about one city. State the real service area and the lanes you run regularly. Building location pages for towns where you have no facility usually makes things worse — we deal with that under our services.
How long before any of this shows up?
Classic measures typically show first movement four to eight weeks after work begins. This layer also depends on model updates that nobody outside can schedule, so plan in quarters and read direction rather than single values.
Our competitors are load boards. Is there any point at all?
Yes, but narrower than a vendor will tell you. You will not displace a platform on generic terms. You can be the source named for the specific things you do — a lane, an equipment type, a commodity, a certification — and those questions end in a phone call rather than a browse.
None of this replaces the work that was already worth doing. Pages that state facts clearly rank well and get cited; pages full of adjectives do neither. What is new is that a machine reads your industry on a shipper's behalf, mostly out of sources you did not write, and the only honest way to track it is an estimate labelled as an estimate every time you show it.
To see the six research views against your own domain, next to the Search Console figures you can verify, set up a Semalt account for your domain and run the assessment once. Read it as orientation, keep the click numbers beside it, and let the freight quote say nothing about either.