AI Keyword Research and Clustering: Risks, Limits and Governance Rules

Language models are fluent, and fluency is persuasive. A chat model asked for “high-volume keywords about solar panels in Izmir” will return a confident list, sometimes with numbers attached. Those numbers do not come from search data unless the tool is connected to a real database. The model is predicting what a plausible answer looks like.

Teams under deadline pressure tend to accept that output, especially when it looks like a spreadsheet. Rules exist so that the convenient answer is not mistaken for the measured one. They also protect the site from slow damage, such as dozens of near-identical articles written from overlapping clusters.

Invented data

Chat models without a data connection can fabricate search volume, cost-per-click and difficulty scores. Even tools that do connect to databases may mix real figures with model estimates without clear labels. Any number in a keyword plan must trace back to a named source.

Wrong intent grouping

Embeddings group words by meaning, not by what Google shows. “CRM software” and “what is CRM” are semantically close, yet one has commercial intent and the other informational. Merging them leads to a page that tries to sell and explain at once, and often ranks for neither query.

Cannibalisation

If clustering runs again each quarter without checking existing pages, it can recommend new articles for topics the site already covers. Two pages targeting the same query split signals and confuse search engines about which one to show.

Data leakage

Pasting a client’s full Search Console export, internal sales notes or unreleased product names into a consumer chat tool may breach confidentiality agreements. Business and enterprise plans from major providers usually offer settings that exclude data from training, but someone has to check and choose them.

Keep the rule set short enough that people actually follow it. Five to eight rules, written on one page, work better than a long policy nobody reads.

• Source rule: volume, difficulty and CPC come only from named tools such as Search Console, Keyword Planner, Semrush or Ahrefs, never from a chat answer.
• Review rule: every cluster mapped to a page gets a live SERP check on its top three terms by a person.
• Inventory rule: before any new page brief, search the site and the content inventory for existing coverage.
• Data rule: client data goes only into approved tools with training opt-out confirmed.
• Prompt rule: keep labelling prompts and intent definitions in a shared file so results stay comparable.
• Ownership rule: one named person approves the final keyword-to-page map.
• Logging rule: record the date, model and settings used for each clustering run.

Use a quick, repeatable test. Take the two or three highest-volume terms in the cluster and search them in a clean browser session for the target country. Compare the top ten results. If they share several of the same URLs, the cluster holds. If one term returns product pages and another returns guides, split it.

For larger projects, SERP overlap tools automate this, but they cost credits. A sensible compromise is to automate checks for the top clusters by value and do manual spot checks on a sample of the rest.

AI clustering does not know your margins, your stock or your sales cycle. A cluster with high volume may be worthless if the products it relates to are out of range or unprofitable. Prioritisation stays a business decision.

Models also reflect the data they were trained on, which skews toward English and toward large markets. For Turkish, Azerbaijani or regional queries, expect weaker labels and check more of them by hand. Local slang, spelling variants and mixed-language queries are common blind spots.

• No real search data without a connected source.
• Weaker accuracy in smaller languages and niche B2B terms.
• No knowledge of business value, seasonality in your market or competitors’ plans.
• Results can change between model versions, so old and new runs may not match.

Ownership works best with one person, not a committee. In an in-house team that is usually the SEO lead; in an agency it is the strategist on the account. That person keeps the rule sheet current, approves exceptions and checks a sample of work each month.

Review the rules every quarter. New tools appear, model behaviour changes and the team learns where errors actually happen. Remove rules nobody breaks and tighten the ones that keep failing. A rule sheet that changes with the work stays useful; one written once and filed away is ignored within weeks.