Keyword-to-Prompt Translation for Search Advertisers Entering AI Channels

Search volume is set to drop 25% by 2026, according to Gartner, and that traffic isn't pausing in some in-between state. It's moving straight into conversational AI. ChatGPT handles 2.5 billion queries a day now, Google's AI Overviews touch nearly half of all Google searches, and a growing share of pre-purchase research starts inside a chat window instead of a search box. US ad spend in AI search is on track to run from roughly $1 billion in 2025 to $25.9 billion by 2029. Advertisers are chasing that number with budget. Almost none of them have changed their strategy to match it.

How a prompt is structurally different from a keyword

A keyword is a fragment, and a stingy one at that. "CRM software pricing" tells an ad platform almost nothing on its own; the platform has to guess at budget, company size, and timeline, filling gaps with statistical inference built from millions of other people who typed something similar.

A prompt does the opposite. It specifies. Semrush's 2026 AI Search Trends report gives an example worth sitting with: "What's the best CRM for a 50-person marketing agency with Salesforce integration that costs under $150 per user monthly?" One sentence carries a use case, a company size, an integration requirement, and a price ceiling, all stated by the user, none of it inferred by a platform working backward from a fragment.

The gap runs deeper than length. Keywords are noun phrases. Prompts are questions, commands, hypotheticals. Google reported that searches beginning with "tell me about…" rose 70% year-over-year in 2025, which says something about how people talk to a search box now, not just what they type into one. In a multi-turn chat session there isn't even a single query to read in isolation; the third message in a thread only makes sense next to the first two, so intent builds cumulatively instead of arriving all at once.

None of this changes what the person actually wants, and that's worth sitting with. Someone asking about CRM pricing for a 50-person agency has the same underlying need as someone who typed "CRM software pricing" into Google a decade ago. What changed is the surface the intent lands on. A targeting system built to read fragments has no mechanism for reading a specification.

What keyword match types were actually doing — and why their logic breaks in conversational environments

Match types were a workaround, not a solution. Broad, phrase, and exact exist because a keyword alone is too ambiguous to act on. Each one is a different bet on how much latitude to hand the platform for guessing at intent the keyword itself never states.

Advertisers have been admitting this for years, mostly with their wallets. Exact match has lost close to 10 percentage points of spend share since 2022, and broad match is now the dominant match type by budget across the industry, advertisers conceding, in aggregate, that rigid string-matching leaves real intent on the table. Phrase match is the interesting holdout. Despite the overall drift toward broad, many advertisers retain it for the structure it provides. Advertisers still want structure; they just stopped wanting exact-match rigidity a while back.

Google's own infrastructure gives this away, too. AI-powered bidding now drives 78% of all Google Ads spend, meaning Smart Bidding is already making the kind of intent inferences keyword match types were never built to make. The industry had been drifting toward signal-based targeting for years before any of this, with platforms like Thrad, a programmatic ad infrastructure built for AI chat interfaces, emerging specifically to serve that gap. It just hadn't hit a channel that forced the issue.

In an LLM environment there's no keyword to match at all. The input is a full sentence, and that sentence might not contain a single word from an advertiser's keyword list while still expressing precise, high-value intent. Multi-turn context makes the mismatch worse, since a match type evaluates one query at a time with no way to account for intent that develops across a conversation. Placement is different too: the ad isn't sitting in a distinct sponsored slot next to organic results, it's woven into the generated response itself. When mobile search took off, Google pushed advertisers into Enhanced Campaigns because desktop-built structures no longer matched how people searched. Search Engine Land has examined the keyword-to-prompt shift as a fundamental structural change in how targeting logic needs to work.

The intent architecture inside a conversational prompt

Every prompt does several jobs at once. Reading them apart, rather than treating a prompt as one flat query, is the actual skill this transition demands.

Start with primary intent: is the person discovering something, comparing options, validating a decision half-made already, buying, troubleshooting? Layered on top is the constraint set, budget, geography, timeline, the systems something needs to plug into, team size. A keyword almost never states any of this. A prompt usually does.

Phrasing alone signals funnel stage. "What is…" and "explain…" sit at the awareness end, someone still forming the category in their head. "Best X for Y" or "compare X vs. Y" sits in consideration, someone narrowing a real shortlist. "How do I get started with…" or "pricing for…" sits near purchase. "Tell me about…" tends to run early-stage but with a specific category already fixed in mind, closer to a warm lead than a cold one.

There's a persona layer, too, which prompts hand over almost for free. "As a head of marketing at a 50-person agency" gives you role, seniority, and company size in eight words, information a keyword search never surfaces directly. Emotional signals ride along on top: urgency, frustration, comparison anxiety. These don't just sort an ad into a category. They shape how the ad ought to be worded.

Multi-turn conversation stacks one more layer on top of that: an accumulating intent file. Systems in these environments increasingly draw on prior turns and stored preferences within a session, building a continuous read of what someone wants instead of treating each message as a cold start. Google's natural-language processing has advanced to the point where targeting logic needs to track topics and the relationships between them rather than string patterns alone. A keyword maps to a category. A prompt maps to a person, a funnel stage, a constraint set, and a decision context, all at once, and reading all four off one sentence is most of the job now.

What carries over from search — the parts of keyword expertise that still apply

None of this means starting over. Intent categorization, the core discipline of mapping a need to a solution, transfers directly. It's the same skill whether you're grouping keywords into ad groups or grouping prompts into intent clusters; only the input format changed underneath it.

Negative targeting carries over too. Years spent building exclusion lists to keep irrelevant or low-converting traffic out of a campaign apply just as well in prompt environments, where knowing which conversational contexts are off-brand still matters. The awareness, consideration, purchase framework still holds. It just gets read off prompt phrasing now instead of keyword type.

Bid modifiers have a clear analogue. Search advertisers already weight bids by device, location, time of day; in AI channels that same instinct gets redirected toward intent strength, constraint specificity, and funnel stage.

The single highest-return bridging tactic is question-based keyword mining. Pulling the exact questions a target audience already asks captures the linguistic register these systems are built to parse in the first place. The Related Searches and People Also Ask feature now appears in 85% of SERPs, per SE Ranking's 2025 research. Building a real library of these questions, rather than guessing at them from a desk, is the most direct path from an existing keyword list to something a prompt-targeting system can use.

Ad copy discipline hasn't gone anywhere either. Relevance, specificity, a clear call to action still matter no matter the format. What changed is how much targeting context arrives already known, before the ad has to earn its relevance the hard way. Measurement instincts carry over too, with one caveat: conversion tracking, ROAS, and CPA are still the right benchmarks to reach for, and OpenAI has started rolling out conversion tracking tools for advertisers, which makes this more workable than it was even a year ago. The tooling is still young, though, and it shows.

What must be rebuilt: the translation mechanics from keyword list to prompt-targeting logic

Start by reclassifying existing keywords by intent, not match type. "CRM pricing," "CRM cost," and "how much does CRM software cost" are three different strings expressing the same underlying thing: a cost-evaluation question from someone mid-funnel. Group by that intent, not by the syntax of the search term, and each group becomes a seed for a prompt-targeting topic instead of a keyword list.

From there, expand each intent cluster into the actual questions someone would ask an AI assistant to satisfy it. Tools like People Also Ask, "tell me about…" expansions, and real customer interview transcripts surface genuine phrasing rather than guessed phrasing. This question set becomes the vocabulary a prompt-matching system is actually built to recognize.

Constraint signals need to sit on top of that as targeting qualifiers. Figure out which constraint dimensions matter for the product (budget thresholds, company size, required integrations) and weight targeting logic toward prompts where those constraints line up with the product's sweet spot. A prompt specifying "$150 per user per month" and "Salesforce integration" is a far higher-confidence signal than the bare phrase "CRM software," and it should be treated as one.

Funnel stage needs to map to message variant next. Awareness-stage prompts, the "what is" and "explain" queries, want educational copy with a low-commitment CTA. Consideration-stage prompts, the "best X for Y" and "compare" queries, want differentiation-forward copy. Near-purchase prompts, the "pricing" and "how to get started" queries, want a direct offer or a trial CTA. This replaces the old keyword-to-ad-group structure with something closer to a prompt-intent-to-message-variant structure.

Negative logic needs rebuilding, not deleting. Research-only framing, competitor mentions used approvingly, purely academic or informational phrasing: all of these are patterns worth excluding, but the exclusion has to work at the level of topic and intent now, since there's no fixed string left to exclude against.

Last is the conceptual shift that matters most: the unit of account has changed. In keyword search you're targeting a string. In AI channels you're targeting a topic cluster defined by intent and constraint dimensions, and platform targeting in these environments already runs on conversational intent, topic, and publisher category. The advertiser's job now is deciding which topic-intent combinations actually map to the product.

How bidding and measurement must adapt when the auction is inside a generated response

The auction runs differently under the hood. Research into retrieval-augmented generation systems has looked at frameworks where an ad's placement in a generated response depends on the retriever's relevance score and the advertiser's bid together, not bid alone and not relevance alone. A higher bid can win better placement, but only inside whatever the retrieval system already judges relevant enough to surface.

Click-through rate gets harder to pin down here. An LLM can fold an ad into a response differently depending on the exact query and context around it, so CTR ends up more variable than on a fixed search results page, where position stays stable and comparable across queries, per analysis published through SIGECOM. That variability makes broader creative testing worth more than it used to be, and it means building real-time feedback loops rather than leaning on a single historical CTR number.

The programmatic layer looks different too. Buying through OpenRTB 2.6 in these environments runs on conversational intent, topic, and publisher category, with no cookies involved anywhere: a real departure from how programmatic buying has worked for the last decade. Early CPM data gives a sense of where pricing is settling. ChatGPT placements are running around $60 CPM; early Perplexity placements in 2024 and 2025 topped $50 CPM. Both sit well above typical display benchmarks, and that premium reflects the quality of intent behind each impression, not inflated demand for the format itself.

Early performance numbers give advertisers something to calibrate against, even without a guarantee attached. Microsoft reports Copilot ads delivering 73% higher click-through rates and 16% stronger conversion rates than traditional search, with customer journeys running 33% shorter. Google's AI Max reporting shows an average 7% lift in conversions at similar CPA and ROAS, and campaigns moving off heavy exact- and phrase-match setups have seen conversion lifts as high as 27%. Those numbers won't hold for every advertiser in every category, but they point the same direction: intent quality, not just reach, is what improves when targeting gets built around prompts instead of keywords.

Measurement infrastructure is still playing catch-up. OpenAI's conversion tracking rollout is a start, but attribution across a multi-turn session, working out which conversational turn actually drove the conversion, remains unsolved across platforms right now. Advertisers are better off building their own measurement frameworks and accepting the gaps than waiting for a finished system to arrive. The metric worth adopting in the meantime is cost-per-qualified-intent-moment rather than cost-per-click: a bid should reflect what it's worth to reach someone who already told you their budget and their integration requirements, not just what it costs to win a click from them.

Where to start: which AI channels are live for paid placements and what each requires of advertisers

Google's AI Overviews and AI Mode are the easiest entry point for a search advertiser, mostly because the account structure and bidding tools underneath are already familiar. Ad coverage inside AI-generated results has grown fast, from about 5.17% of AI results in early 2025 to 25.5% now, and Shopping ads with Direct integration already appear inside generated answers themselves rather than beside them. For an advertiser already running Google Ads, this is the lowest-friction place to start testing prompt-based targeting against a live auction, building on the reclassification work described above rather than porting an old keyword list over unchanged.