AI Keyword Research: Turn Search and Grounding Data Into Topics

AI keyword research combines conventional query evidence with grounding and prompt observations, but it must not convert every conversational phrase into a fake search-volume keyword. The output is a topic-and-task map with known demand, observed AI behavior and clearly labelled hypotheses.

Use four evidence streams without blending them

Google’s generative AI performance reports show impressions, pages, countries, devices and dates for eligible sites in the rollout. They do not expose grounding queries. Bing’s AI Performance report does provide sampled grounding-query phrases for supported Microsoft experiences. Preserve each product’s scope and unit; neither export is conventional search volume.

Begin with a business problem

Write the customer decision in plain language: choose a provider, fix a technical issue, compare approaches, understand a change or complete a task. Add the product, audience, geography and constraint. A topic is valuable when helping that decision can create trust or a useful next step, not merely when an AI tool generates many variants.

Use our topic-cluster method to define one primary job per page.

Collect measured search evidence

Export Search Console query-page pairs with clicks, impressions, position and dates. Add internal site-search terms, sales questions and support tickets. Keep zero-click queries; they can reveal language even when your page did not earn a visit. Protect privacy by aggregating or redacting customer data.

Keyword-tool volumes are estimates. Store vendor, market, match logic and retrieval date. Round numbers appropriately rather than displaying false precision.

Collect AI and grounding observations

Create a fixed prompt set from real jobs, not a list designed to mention your brand. For each run, save exact wording, platform state, locale, answer, cited URLs and any reported grounding query. If the platform does not expose a grounding query, mark it unavailable, do not reverse-engineer a confident keyword from the prose.

Separate a citation observation from demand. A page cited once for a niche prompt may be useful evidence of source fit, but it does not establish a high-volume opportunity.

Cluster by task and evidence

Group phrases when the same page could satisfy the same decision with the same evidence. Split when the audience, method, risk or expected artifact changes. A “what is” query, migration checklist and buyer comparison often belong to different pages even if they share a noun.

For each proposed URL, name the information gain: original measurement, tested workflow, annotated template, comparison method, expert record or primary-source synthesis. If you cannot name the gain, improve an existing page instead of adding another URL.

Prioritize with transparent scores

Use a simple decision score rather than a proprietary “AI potential” number. Rate business fit, evidence strength, unmet usefulness, production effort and maintenance burden on defined scales. Keep every input visible. A lower-volume topic with first-party data and strong product fit may outrank a broad term dominated by generic summaries.

Review the overlap against your current inventory, including the core AI SEO guide and related measurement pages.

Validate after publication

Track index status, query-page alignment, citations, referrals and conversion outcomes through a fixed window. Annotate edits. If two pages split the same query set, decide whether distinct conversions justify both. If a page earns impressions but no engagement, revisit its promise and usefulness before generating more content.

Where SISTRIX Prompt Research fits

SISTRIX’s October 5 Q3 recap documents Prompt Research access through API functions ai.topicresearch.overview, ai.topicresearch.list and ai.topicresearch.topic, plus its MCP Server. This is a quarterly publication date, not proof that every feature launched that day. SISTRIX’s question and topic counts describe its vendor dataset, not Google search volume.

Use the overview to frame a topic, the list to identify candidate questions, and topic detail to inspect the proposed brief. Check account access, dataset market and language coverage before interpreting an export. No SISTRIX API export was obtained for this update.

A fictional brief, with the evidence columns kept separate

Suppose a topic tool suggests “How do I stop translated Shopify descriptions becoming stale?” That is a candidate reader job. If your own Search Console export shows impressions for a related phrase, retain the actual query, page, period and count in a separate field. Without that export, mark measured site demand unavailable rather than assigning a made-up number.

Then check the current library. If one page already explains the same freshness decision, update it. If the new requirement is a different decision with different evidence, it may deserve a new page. For example, Shopify’s weekly translation change needs a source-to-locale timing explanation, while AI-shopping price pairs need variant and currency checks. Both concern commerce, but they do not answer the same job.

Before accepting an automatically generated brief, ask it to name the claim requiring a source, the worked example the reader will receive, and the existing URL it might duplicate. Reject generic briefs that cannot answer those questions. A larger prompt dataset can improve discovery coverage; it does not justify producing more pages than the publication can maintain.