Using AI for keyword research

What it does well

Generating candidates. Ask what somebody hiring your trade would want to know and you get a long list quickly, including phrasings you would never produce yourself because you are too close to the subject.

Grouping by meaning. Given fifty phrases, sorting the ones that are the same question into a single group is exactly the work that takes an afternoon by hand and is easy to verify once done.

And sorting by intent. Someone comparing options, someone ready to hire and someone asking an answer engine all need different pages. Our page on search intent covers why that split matters most.

Where it invents things

Search volumes. A model has no access to query counts, so a monthly figure offered in a list is a guess shaped like a statistic. It will look plausible, it will be specific, and it will be made up.

Difficulty scores have the same problem, with the added complication that every tool defines difficulty differently. A number with no method behind it is worse than no number, because it feels like evidence.

Treat any figure as something to verify elsewhere or delete. This is the same rule our page on citable claims applies to everything else you would publish.

A workable process

Generate wide, then cut hard. Ask for questions, phrasings, problems and objections across your whole trade, then remove everything that is not a question a paying customer would actually have.

Group what remains into one page per meaning. This is the step that protects you from the doorway pattern, because it is where ten phrasings of one question become one page instead of ten.

• Collect candidates from the tool and from your own inbox
• Group everything that means the same thing
• Name the single question each group answers
• Mark which groups lead to paying work
• Check volumes in a tool with real data, or skip volumes entirely
• Decide the pages you will not build
• Write the page titles before writing any page

Your own records beat any tool

The questions in your inbox came from real people who were close to hiring somebody. That is better evidence of demand than any estimate, and nobody else in your market has the same list.

Search suggestions are the next best source, because they come from real queries. Typing the start of a question about your trade and reading the completions costs nothing and produces genuine phrasing.

Then confirm against reality. If a subject looks promising but you have never once been asked about it in years of trading, that is information worth trusting over a list.

Turning a list into pages

One page per meaning, never one per phrasing. A page built for each variant of the same question competes with its siblings, which our page on competing pages describes in detail.

Give each page a title that names the question in plain words, and put the answer near the top. The list is only useful once it has become a plan, and a plan is a set of titles with a purpose each.

Then order the build by commercial value rather than by ease. The pages that bring in work should exist before the pages that are interesting to write, however tempting the reverse feels.

Local and trade phrasing

Models tend toward national and generic wording. Real customers use the town, the neighborhood, the slang name for the job, and the brand of the thing that broke. Those specifics rarely appear in a generated list.

Add them yourself from how customers actually speak on the phone. A page that uses the words a local uses reads as local, which is a signal no amount of general optimization replaces.

Our page on local search basics covers the rest of that work, including the parts that happen off your own site entirely.

What to do with the leftovers

Most of a generated list will not become pages, and that is the correct outcome. Keep the discarded items with a note saying why, so the same subject does not get reconsidered every six months.

Some will become sections inside existing pages rather than pages of their own. That is usually the right home for a question that is real but small, and it keeps the site from spreading thin.

A few will be worth revisiting after your services change. Keeping the list is cheap. Rebuilding the thinking behind it a year later is not, and the reasoning is the valuable part.