Sameer Khan
The Death of Keywords: Why Conversational Commerce is Taking Over
Short answer: Keywords are not disappearing — organic search still carries the overwhelming majority of commerce traffic. What is changing is the shape of the query. Shoppers increasingly describe a situation in full sentences rather than compressing it into two or three searchable terms, and the systems answering them match on meaning instead of string overlap. That breaks merchandising built around keyword matching, even while keyword search itself remains large.
The practical shift is from optimising for terms to optimising for situations.
What actually changed
For twenty years shoppers did the translation work. Someone who wanted comfortable shoes for their first 5K with flat feet typed “running shoes flat feet” — discarding the race, the experience level and the budget, because the search box could not use them. Retailers then optimised for the compressed version, and everyone accepted the information loss as the cost of using a search engine.
Assistants removed that constraint. The same shopper now writes the whole situation out, and every detail they used to discard is a constraint the system can act on. More signal reaches the retrieval layer than a keyword ever carried.
The uncomfortable implication: the extra detail is precisely what disqualifies you. A keyword search for “running shoes flat feet” returned everything arguably relevant. A described situation returns only products that demonstrably match all the stated constraints. Vagueness used to keep you in the running. Now it removes you.
Why keyword tactics stop transferring
Keyword optimisation rested on matching — put the term on the page, in the title, in the anchor text, and rank for it. Semantic retrieval does not work that way. A product qualifies for “shoes for a beginner runner with flat feet” when its attributes establish arch support, cushioning and an entry-level profile, whether or not that phrase appears anywhere on the page.
Which means the classic response — write a page targeting the long-tail phrase — largely misses. The phrasings are effectively unlimited and vary per shopper. You cannot enumerate them. What you can do is make the underlying facts explicit enough that any phrasing resolves to your product, which is the entire argument for machine-readable product attributes.
The four things this breaks
Keyword-led content planning. Building pages per search term produces thin near-duplicates that answer no situation well. It is also the pattern Google's spam policies now describe as scaled content abuse — many pages generated without adding value. The volume approach has gone from ineffective to actively risky.
Copy that implies rather than states. “Designed for all-day comfort” cannot be matched against a constraint. “Cushioned midsole, 10mm drop, suited to neutral and flat arches” can. Persuasive copy still matters for the human; only explicit attributes matter for the system deciding whether the human ever sees you.
Ranking as a success metric. Position in a list is not the outcome anymore. Inclusion in an answer is, and the two correlate weakly — fewer than 10% of sources cited by ChatGPT, Gemini and Copilot rank in Google's organic top 10 for the same query.
Attribution. A conversation that ends with your brand name and a direct visit records as direct traffic. The query that decided the purchase never touches your analytics.
What to do instead
Inventory the situations, not the terms. Your support tickets, sales calls and reviews already contain how customers describe their problem in their own words. That language is the new keyword research, and you already own it.
Make every constraint explicit and structured. For each situation, ask which stated facts would be needed to qualify. If a fact lives only in a product photo or a paragraph of prose, it does not exist for matching purposes.
Answer at the level of the situation. One thorough page covering “choosing running shoes for flat feet” — with specifications, trade-offs and honest disqualifications — outperforms nine thin pages targeting nine phrasings, and does so under both search and assistant retrieval.
Say who it is not for. Assistants are matching constraints, so a stated limitation prevents a bad recommendation and makes your remaining claims more credible. This is one of the few places where narrowing your claims widens your reach.
Keep the nuance
It would be convenient to declare search finished. The numbers do not support it: organic search remains a very large majority of commerce discovery, and assistant-driven traffic, while growing fast, is still a small fraction of total internet traffic. Anyone telling you to abandon SEO is overselling.
The accurate version is narrower and more useful. A second discovery layer has appeared alongside search, it rewards different signals, and the shoppers using it arrive further along in their decision. You are not replacing one practice with another — you are adding a second scoreboard, which is the argument laid out in the shift from SEO to GEO.
The bottom line
Keywords are not dead. Keyword compression is — shoppers no longer flatten their needs to fit a search box, and the systems answering them reward products whose facts are explicit enough to satisfy a fully stated request. Make the facts explicit and you become matchable across every phrasing you could never have enumerated, which is exactly how AI assistants choose what to recommend.
Want to see which situations your products already win? Book a demo with agentShop for a read on your share of recommendation across ChatGPT, Gemini and Google AI Mode.
