Semrush’s “Keyword Strategy in SEO: What It Is & How to Create One” updates the familiar keyword-planning workflow for a world that includes ChatGPT, Google AI Mode, and AI Overviews. It sensibly argues that competitor rankings and search-volume databases describe only what has already happened in traditional search. The guide also improves on basic keyword research by asking marketers to consider business goals, resources, formats, competition, and the likelihood that a result will actually receive attention.
Its central shortcut, however, undermines that modernization. Semrush says it will use “keyword” to mean both a traditional search query and an AI prompt. Putting both in the same spreadsheet does not make them strategically equivalent.
A keyword is an aggregated, repeatable string with at least some observable relationship to search demand. An AI prompt is often one turn in a longer, personalized exchange. Its meaning may depend on previous messages, uploaded files, location, account history, model version, system instructions, and the assistant’s own follow-up questions. Two people can express the same need with different prompts; identical prompts can produce different answers; and one person’s conversation can change direction without ever repeating the phrase a marketer decided to “target.”
The guide acknowledges this mismatch by telling readers to enter “N/A” for prompt volume and difficulty. It then makes a much stronger claim: if a prompt appears in the tool, “you know there’s demand.” That conclusion needs evidence the article does not provide. A monitored prompt may be drawn from observed behavior, generated from a topic, selected by a vendor, or included because a competitor happens to appear for it. Those sources have different implications. Presence in a tracking database proves that the prompt is being tracked, not that a meaningful customer segment frequently asks it.
This matters because prompt-gap analysis can reproduce the same weakness the article identifies in old-fashioned competitor analysis. If marketers pursue questions mainly because rivals appear for them in a vendor’s prompt set, they are still following competitors through a historical sample. The interface has changed, but the strategy remains reactive. Worse, a curated benchmark can encourage many brands to produce similar pages for the same visible prompts, making the content landscape more homogeneous while leaving unmeasured customer questions untouched.
The recommended manual analysis is fragile too. Typing a prompt into ChatGPT or Google AI Mode and recording the brands, citations, and formats in one response is not a view of “the AI landscape.” It is one observation under one set of conditions. A defensible test would specify the model, date, account state, geography, prompt variants, number of repetitions, and criteria for inclusion. Without that protocol, a marketer can mistake normal output variation for a competitive gain or loss.
Even reliable visibility data would not complete the strategy. A citation can support a brand, mention it neutrally, or use it as a negative example. A recommendation can reach a user who was never likely to buy. An AI answer may satisfy the customer without generating a visit that analytics can connect to revenue. Share of mentions is therefore no more a business outcome than average ranking position. It needs to be paired with recommendation context, qualified demand, assisted conversion evidence, acquisition cost, and incrementality.
The guide is right to organize work around broader topics and genuine audience needs rather than isolated phrases. But its suggestion that topic clusters help because they mimic how AI answers cover subjects reverses the logic. A company should build connected evidence because customers need coherent, trustworthy answers and because the business has something useful to contribute—not because a model’s current response format happens to look similar.
The addendum is to keep prompts out of the keyword column. Use them as probes into customer decisions: group them by underlying job, uncertainty, comparison, constraint, and evidence need. Validate those needs through support conversations, sales calls, product research, site behavior, and commercial results. Then publish material or structured product information that resolves the real uncertainty better than competitors do. Keywords can estimate a market for repeated expressions. Prompts can reveal possibilities. Confusing the two turns the absence of data into the appearance of strategy.