Computational Marketing
FeaturesLong read

"How many keywords for SEO" is the wrong unit: a page owns a question, and you can test the grouping with your own data

Group queries by shared search results and AI-answer coverage.

· 8 min read
Cover illustration for “"How many keywords for SEO" is the wrong unit: a page owns a question, and you can test the grouping with your own data”
Features · November 20, 2026 · 8 min read · 1,871 words

A page should own a question. The useful answer to “how many keywords for SEO” comes from testing which queries one page can satisfy. Shared search results reveal the first grouping, while AI-answer citations and follow-up questions test its coverage. Keyword count is what falls out of that work.

Why the keyword count per page is an output, not an input

“How many keywords per page?” hides an assumption that count is the right unit of planning. It treats the number as something to choose before deciding what the page should answer. That gets the order wrong. Two pages can both target eight terms, yet only one may group queries that genuinely share an intent. The other has only a quota.

The older one-keyword-per-page approach made this flaw visible through fragmented URL structures and pages competing with each other. Giving each page several keywords does not solve that problem if the groups still have no clear basis. First, determine which searches belong together. Once that is settled, the number of terms describes the size of the group.

A useful brief names the primary query and its verified secondary variants, so the author knows both the central question and its scope. The number can still appear in the brief. It simply records a decision made from evidence.

What actually determines whether two queries belong on the same page

Shared words are an unreliable guide to shared intent. Consider “seo audit” and “seo audit tool,” which look close enough to combine at a glance. Their search results point in different directions. The first brings up agency content and editorial guides, while the second brings up lists that compare software.

Shared ranking URLs provide a more direct signal because they show which pages Google selects for each query. If substantially the same pages rank for both, the results indicate that one page can satisfy both searches. If the results differ, there is a stronger case for separate pages.

How much overlap is enough to act on? A soft threshold of one to two shared URLs produces larger clusters, but it also allows more false merges. A moderate threshold of three to four shared URLs is the practical default for established sites with real link equity. The threshold shapes the outcome. Using a tool’s default without knowing why it was chosen can put different needs into a cluster that looks tidy.

There is a limit to what this measurement can settle. Very low-volume queries can have thin or unstable search-result data, making overlap a less dependable signal. Those queries need manual intent review before they are put in a group. Semantic similarity can help break a tie in that situation. It can help with the decision, but shared results should still guide it when they are usable.

A runnable method to group queries by shared results and shared AI-answer citations

To put that test into practice, start with the raw query set. Export Google Search Console queries and impressions for the domain over a recent multi-month window, then include seed terms, long-tail variations, questions, and competitor keywords. Do not filter for volume yet. Early filtering would narrow the evidence before the groups have taken shape.

Next, get the top ten ranking URLs for each query through a SERP API. Compare every pair of queries and count their shared URLs to build an overlap matrix. Apply the chosen threshold to form clusters. Each cluster then needs one intent label, such as informational, commercial, or transactional, and one target URL. If a cluster needs two labels, split it before moving into page planning.

Existing pages come before new content commissions in this process. Match each cluster to a current URL first, then identify the groups that genuinely lack a home. The working output is a three-column sheet containing the cluster, target URL, and action, with the action set to keep, merge, redirect, or create. A long list of queries can shrink to far fewer clusters. That compression helps expose how much of the apparent content workload came from fragmentation.

The second pass tests something that shared ranking URLs alone cannot show. Google’s AI Mode launch announcements describe query fan-out, which issues multiple related searches concurrently across subtopics and multiple data sources. The head query is therefore only the starting point. A candidate page needs to answer the related questions an answer engine generates. This adds a coverage test to the intent test.

For each cluster’s primary query, enumerate the fan-out questions generated in AI answer engines and inspect whether the candidate page answers each explicitly. Check the shared citations as a second dimension of the grouping. Where does the page answer the question directly, and where does it leave the reader to infer an answer? Record the gaps, including questions covered only by a passing mention. Clusters that satisfy both the SERP-overlap test and the test of coverage for related questions generated by answer engines become the highest-priority pages to build or strengthen.

An answer-first opening addresses the primary query, while explicit definitions, named entities, concrete verifiable claims, and tables help make important passages understandable when extracted on their own. Keyword density cannot do that work. Recency also belongs in the process, because AI citation rates can drop when content ages beyond roughly three months without an update. The cluster map therefore needs to support maintenance as well as initial production.

A machine-readable page-to-question map and a check for pages competing with each other

Before content is written or combined, each approved cluster needs a record that is kept and tied to a specific URL. The page-to-question map gives that decision a testable form through the fields url, primary_query, cluster_queries[], intent, ai_fanout_questions[], owner_page, and action. Each field has a job. Together, they identify the page, the queries it should answer, its intent, its AI-question coverage, and the next approved change.

A cluster marked “create” becomes a content brief, while “merge” becomes a consolidation ticket and “redirect” becomes a redirect task. “Keep” records that the existing assignment stands. This turns a grouping exercise into a production queue without assuming that every cluster needs fresh content. It also makes unresolved ownership visible before time is spent writing.

Keyword cannibalization occurs when two or more URLs on the same domain compete for the same query and intent, splitting impressions and swapping positions. Google Search Console provides a way to inspect this by filtering for one query and opening the Pages tab. Multiple URLs are a signal to investigate. The shared intent and competing performance distinguish the problem from a simple count of pages appearing for a term.

For a programmatic check, pull per-URL impressions and clicks for each query through the GSC API. The resulting query-to-URL matrix shows where the domain has overlapping exposure. Flag a query appearing for two or more distinct URLs with meaningful clicks as a cannibalization candidate. Then compare those URLs with the page-to-question map to decide whether they belong to one cluster or serve separate intents. A flag starts the decision. It does not settle it.

The distinction prevents automation from flattening every overlap into a merge. When pages genuinely compete for one intent, the action record gives consolidation or redirection a named destination. When they serve different needs, the map preserves that separation. The AI-citation check then tests whether the chosen owner answers the cluster’s fan-out questions without relying on competing pages to complete the answer. Ownership requires coverage as well as URL housekeeping.

A checklist an agent could execute

The workflow can be expressed as a sequence of structured outputs, with a human approval gate before changes ship. For the first step, an agent pulls all available GSC queries and impressions for the domain over the last 90 days. Its output is a flat CSV with query, impressions, clicks, and average position. That is the starting record for the run. It keeps the later cluster and page decisions tied to an explicit query set.

The second step retrieves the top ten ranking URLs for each query through a SERP API. Its output is the query-to-URL matrix used to compare search results. This matrix represents external ranking overlap, with competition among pages within the domain assessed separately. Keeping those purposes distinct is important. The same query can share ranking results with another query while also exposing an internal ownership conflict later in the process.

The third step compares each pair of queries and counts the URLs they share. It groups queries that meet the chosen threshold, using three to four shared URLs as the default for most established sites. The output is a cluster membership table. Low-volume queries with unstable results need manual intent review. This is where the process separates a measurable grouping from a case that still needs judgment.

The fourth step assigns an informational, commercial, or transactional intent and a single primary query to each cluster. It produces a cluster intent table. Any group requiring two labels is flagged for a manual split decision before page assignment proceeds. The flag should stay in view in the action list. Otherwise, uncertainty in the grouping can disappear beneath an apparently complete content plan.

The fifth step matches each approved cluster to an existing URL before marking any gap as “create.” Its output is the page-to-question map, initially containing url, primary_query, cluster_queries[], intent, and action. The full map also lists the owner page and, once they have been checked, AI fan-out questions. Existing coverage gets the first look. A new-content brief follows only when the matching step leaves a genuine gap.

The sixth step pulls per-URL impression and click data for each query from the GSC API. It flags queries appearing for two or more URLs with meaningful clicks and produces conflict pairs for review. Each pair receives an action field for merge, redirect, or keep with canonical. These are proposed actions. The competing pages still need to be judged against their intent and the map’s chosen owner.

The seventh step runs the AI-coverage check for the head query of every cluster. The agent enumerates the fan-out questions generated in AI answer engines and tests whether the target page answers each explicitly. The output is an AI-coverage gap list. A cluster can pass the ranking-overlap test and still fail this check. Keeping the two results separate shows whether the next task is to settle page ownership or fill in missing answers.

The eighth step presents the action list for human approval. A person approves merges, redirects, and new-content briefs before anything ships, with no consolidation or redirect executed without that signal. The agent can assemble evidence and propose changes, but a completed table is not permission to alter the site.

The ninth step measures the result after consolidation or new content goes live. Re-run the AI-citation check across ChatGPT, Claude, Gemini, and Perplexity to assess citation recovery, then schedule the full workflow again, with quarterly as a reasonable default for most sites. The outputs are used in the next run as well as in the report. Human attention stays on the action list, while the underlying records make each proposed change inspectable. The final planning question is no longer how many keywords a page has, but whether the evidence supports the question it owns.

More in Features