Choosing keywords: the method we run on our own site

Most keyword research methods start with a tool and a column of search volumes. For a Quebec SMB, that's backwards. SEO has one job here: bringing in profitable inquiries. So the opening question isn't "which word gets searched most", it's "which service do we want to sell more of, and how does a buyer ask for it?"

This article describes the method we run on our own site, the site of a company selling websites and AI agents on subscription. Not theory: our real keywords, our real numbers, including the disappointing ones. The main keyword of the article you're reading gets 30 searches a month in Canada. We're writing it anyway, and you'll see why.

The whole thing fits in one sentence: in French in Quebec, volumes are structurally thin, so intent and customer value beat volume almost every time.

The short answer, for the busy

Choosing keywords starts with your services, not with a tool. List the offerings that make the most margin, then write down the words a buyer would use to ask for them. Only then check reality: in French in Quebec, volumes are thin, so intent matters more than the search count. Our own example: "refonte site web" (website redesign) draws only about 70 searches a month in Canada, yet an ad click on it sells for US$31.11, which tells you what a project is worth to competing agencies. The free toolkit is enough: the Search Console queries where you already sit on page two, Google's autocomplete predictions, the related-questions block in the results, and Keyword Planner for orders of magnitude. Then build one page per intent, never one page per spelling variant. And ignore keyword density and "LSI keywords": the latter don't exist.

Start with the money, not the tool

Take the last twelve months of revenue and rank your services by margin, not by units sold. The one at the top is your starting point. Then ask a blunt question: if an ideal client opened Google tomorrow morning to find that service, what would they type?

That question beats a tool, because it produces buyer vocabulary instead of seller vocabulary. Companies write "document management solutions"; clients write "scan paper files". The gap between the two is where most SMBs lose their traffic.

The practical test: dig through your inbound emails and contact form submissions. People describe their problem there in their own words. That's your first list, and it costs nothing.

In French in Quebec, volume tells you almost nothing

Here are our own figures, measured in July 2026 for francophone Canada via DataForSEO (Google Ads data, in US dollars). A click on "refonte site web" sells for US$31.11, against roughly 70 searches a month. "Créer un site web": US$16.52 for about 260 searches a month. "Agents IA": US$9.53, also for about 260 searches a month.

Look at that first line. Seventy searches a month is nearly nothing. A volume-first tool would tell you to drop it. But the click price tells the other half of the story: if advertisers will pay thirty dollars for one visit, a redesign project comfortably justifies that bet. A keyword's cost per click is the best free indicator of what a customer is worth in that market. Volume tells you how many people are searching; the auction tells you what each one is worth. In thin markets, only the second number matters. That's also the whole argument behind our piece on choosing between SEO and Google Ads.

Two honest caveats. First, these estimates move a lot. Re-running the same query on 17 August 2026, "refonte site web" came back at US$62.75, double the July figure. On a low-volume keyword the estimate is noisy: keep the order of magnitude, never the decimal. Second, a high cost per click doesn't mean you should buy the click. It means the page answering that query deserves your best work.

The extreme case is this article's own keyword. "Mots clés référencement" shows 30 searches a month in Canada, zero advertiser competition, and no cost per click at all: nobody is bidding (measured 17 August 2026, same source). A purely quantitative reading would say don't write this article. We're writing it because the person asking that question is exactly the person who, six months later, goes looking for an agency.

The free toolkit is enough

We have never paid for a keyword tool subscription for our own site. Four free sources do the job.

The Search Console performance report. It's the only source that talks about you rather than about the market. Switch to the "Queries" dimension and sort by impressions. You're hunting for queries with plenty of impressions and an average position between 8 and 20: Google is already showing you, just not high enough. Those are the cheapest wins, because half the work is done. The metrics available are clicks, impressions, CTR (clicks divided by impressions) and average position, defined by Google as the position of your site's top-ranking result.

Google's autocomplete predictions. Type your service into the search bar and read what Google offers. Per Google's documentation, "autocomplete predictions reflect real searches that have been done on Google", weighted by factors including the language of the query and the location it comes from. So it's a direct read of the real vocabulary, in French, in Quebec. Free, instant, and more trustworthy than a rounded estimate.

The related-questions block. Google officially calls it a "related questions group"; everyone else calls it "People Also Ask". Google itself suggests mining it: its documentation notes that "it may be helpful to pay attention to related search queries when brainstorming topics to write about for your site." Every question there is a potential page title, already phrased by humans.

Google Ads Keyword Planner. Useful for orders of magnitude and for the top-of-page bid range, shown as two columns (low and high). Treat its volumes as magnitudes, not measurements.

One page per intent, not one page per variant

This is the costliest mistake we see at SMBs: building "plumber Laval", "plumber in Laval" and "best plumber Laval" as three separate pages. They cannibalize each other, none accumulates authority, and Google has a name for the practice.

Its spam policies describe "doorway abuse" as when "sites or pages are created to rank for specific, similar search queries", citing among the examples "creating substantially similar pages that are closer to search results than a clearly defined, browseable hierarchy". Three phrasings of the same request are one intent. So: one page, done seriously, that handles it better than anyone else's.

The right split follows what the person wants to accomplish. "How much does a website cost" and "web agency Montreal" are two intents: the first wants a number, the second wants a vendor. Two pages. "Website price" and "cost of a website" are one intent with two phrasings. One page.

What we deliberately ignore

Big generic head terms. "Marketing", "website", "artificial intelligence". The volume flatters, but the intent is unreadable and the competition is global. An SMB chasing those terms is buying frustration.

Keyword density. There is no target percentage. What does exist, in Google's spam policies, is "keyword stuffing", defined as "filling a web page with keywords or numbers in an attempt to manipulate rankings", with the example "repeating the same words or phrases so often that it sounds unnatural". Write to be read.

"LSI keywords". They don't exist. Google's John Mueller said so in 2019 ("There's no such thing as LSI keywords"), then again in January 2023 ("Both have no effect. Anyone who tells you to use LSI keywords is ... still wrong after all these years"). If a vendor is billing you for "LSI optimization", you now know what you're buying.

Where to start this week

Three hours gets you a usable first list. One hour ranking your services by margin and writing, for each, the three phrasings a client would use. One hour in Search Console spotting the queries where you sit between position 8 and 20. One hour typing those phrasings into Google to harvest the autocomplete predictions and related questions. You'll come out with about ten intents, each deserving a page. Then be patient: SEO timelines run in months.

If you'd rather we look at your situation first, that's exactly what our audit covers: which queries you already appear for, which ones are within reach, and which pages are cannibalizing each other. 30 minutes, no commitment.

→ Request an audit