September 20, 2026
Google answers the industry question, not the local one
Eleven Canadian hiring searches, three runs each. Five of six industry searches returned an AI answer every time. All five city searches gave a map pack.
By Lasse Pettersen, independent SEO consultant, Collingwood
On 21 September 2026 I measured eleven Google searches in Canada, the kind somebody types when they want to hire an SEO. Five were city searches, run from an IP address inside each city. Six were industry searches, run nationally. I ran each one three times rather than once, which turned out to matter.
The measurement
Three runs per query, within about two hours.
| Search | AI answer | Map pack |
|---|---|---|
| seo consultant toronto | 0 of 3 | 3 of 3 |
| seo consultant calgary | 0 of 3 | 3 of 3 |
| seo consultant vancouver | 1 of 3 | 3 of 3 |
| seo consultant edmonton | 0 of 3 | 3 of 3 |
| seo consultant montreal | 0 of 3 | 3 of 3 |
| wordpress seo expert | 3 of 3 | 0 of 3 |
| ecommerce seo expert | 3 of 3 | 0 of 3 |
| shopify seo expert | 3 of 3 | 0 of 3 |
| saas seo consultant | 3 of 3 | 0 of 3 |
| law firm seo expert | 3 of 3 | 0 of 3 |
| dental seo expert | 0 of 3 | 0 of 3 |
Two patterns, and they are close to absolute. Every city search returned three map pack slots on every run. Five of the six industry searches returned an AI-written answer on every run. Not one search returned both features on any run.
Why this is not a coincidence
Add a city to a hiring query and you change what is being asked. “WordPress SEO expert” is a question about a category of service. “SEO consultant Toronto” is a question about who, near me, can be hired. Google appears to treat the first as answerable and the second as a matter for the map.
That is consistent with something I measured earlier on a different set of queries: the model writes the answer when the answer is a comparison, and stands back when the searcher needs a person. Naming a city is the clearest signal a searcher can send that they need a person.
The exceptions, which are the useful part
Dental was the reason I ran this three times. On my first pull of the day it returned an AI answer. Two hours later it returned none, on three consecutive runs, showing a knowledge panel and a block of Google reviews instead. A fifth check later the same afternoon returned an AI answer again. Same query, same country, same tool, five observations, three different answers about whether the feature exists.
Vancouver produced an AI answer on one run of three and nothing on the other two, with the same query, the same location and the same tool.
So the conclusion is not “these searches have AI answers”. It is: on a given query, the presence of an AI answer is stable for some and unstable for others, and a single check tells you nothing about which kind you are looking at. If you measured dental once this morning you would have published the wrong finding. I nearly did.
Anybody reporting AI Overview coverage from one pass per query is reporting noise mixed with signal and cannot tell you which is which. Three runs is a crude fix and it is better than one.
What it changes if you are trying to be found
On city queries, the pack is the prize, and the pack is proximity-led. Three slots sit above every organic result, on every run I made. Winning them is a Google Business Profile problem: categories, service area, reviews, consistency. If you have no address in that city you are competing for the organic results underneath, and the right move is to say so rather than invent an office. I wrote up what each of the five city pages actually looks like, including who holds first place.
On industry queries, the answer is written before anybody clicks. Ranking third under a generated answer is worth less than ranking third used to be. The job shifts from being the best result to being the kind of source an answer is assembled from: quotable, specific, and described accurately on sites you do not control. The industry pages note which of them currently carry an answer.
Those are two different jobs. One is profile hygiene and proximity. The other is publishing and being cited. A single plan that treats them as one activity will underperform at both.
The uncomfortable part for consultants
Five of the six industry searches I measured are searches somebody like me would want to rank for, and all five now have a machine-written answer above the first result, explaining what such a person does and what to look for, without sending the reader anywhere in particular.
That is a reasonable thing to do for the searcher and it quietly removes the top of the funnel for the seller. The response is not to write more pages about being an SEO expert. It is to publish things that get cited, which is slower, harder to measure, and the only durable version of this work I can see.
Method, and a correction already
Searches run through the DataForSEO live SERP endpoint on 21 September 2026, desktop, English. City searches used a location inside the named city, industry searches used Canada. The table reports three runs per query at a depth of ten results, within a two hour window.
The correction: my first pass was a single run per query at a depth of twelve, and on that pass dental returned an AI answer, which would have made the headline “six of six”. The three-run measurement replaced it before this was published. I am leaving the account of the error here because the error is the point of the piece.
Positions and features move. Treat all of this as a dated snapshot, not a claim about today. If you re-run these and get something different, I would like to know: lasse@lassepettersen.com.