Google AI Mode Ads Up-Funnel: Keyword Research 2027

Introduction

For twenty years, keyword research followed a simple gravity: find the phrase people type when they are ready to buy, then buy or rank for it. Bottom-funnel terms won; everything else was "awareness" — nice to have, hard to prove, easy to cut.

That gravity has inverted, and Google said so out loud twice in one afternoon.

On October 28, Google's head of search data science told advertisers that people now arrive at Google much earlier in the buying journey, asking long, detailed questions that start a long way from a purchase. The same afternoon, Alphabet's chief business officer told Wall Street that the AI Mode queries driving that shift make Google less money than traditional searches do.

Sandwiched around those statements, Google shipped the commercial plumbing for that behavior: multi-turn ads inside AI Mode follow-up questions, AI Mode ads expanded to ten more countries, home services booking links, and product feeds turned into ad creative. A new Adthena report meanwhile counted more than 175,000 paid placements inside Google's AI answers in the US between June and September.

The queries moved up the funnel, the ads followed them up, and most keyword research is still aimed at the bottom. Here is what changed, what the numbers say, and how to rebuild a keyword strategy that can see the new inventory.

What Changed in the Last Two Weeks

The pieces arrived fast enough that the pattern is easy to miss. The timeline that matters:

October 28 — Google announced multi-turn ads in AI Mode in the United States, so an ad can appear on follow-up queries inside the same conversation. Shashi Thakur, VP and GM of Google Search Ads, confirmed the US rollout.
October 28 — Google said ads in AI Mode are scaling to ten additional countries, including Canada and Australia, with Brazil, Mexico and the United Kingdom planned for Q4 2026.
October 28 — On Alphabet's Q3 call, CBO Philipp Schindler attributed part of the revenue slowdown to "the strong growth in AI Mode queries... because of the lower monetization we were seeing in AI Mode."
November 4 — Google AI Mode added home services booking links, turning an answer into a transaction surface.
November 4 — Reporting confirmed product feeds are becoming the new ad creative in AI Mode: the feed, not the headline, now does the persuading.
November 9 — Coverage of the rollout flagged the measurement problem: AI search activity is nearly invisible in Google Ads search terms reports.

Google is not being subtle. The fastest-growing query type on its surface is conversational, long, research-stage, and now monetizable in more markets than it was three weeks ago.

What Google Actually Said About Those Queries

The most useful primary source of the fortnight is the October 28 episode of Google's own Ads Decoded podcast, where host Ginny Marvin interviewed Shivani Mohan, Google's VP of Data Science and User Experience Research for Search. Her teams interview users one-on-one and study logs of what billions of people do, which makes her statements the closest thing to a specification for where query behavior is heading.

Queries are three times longer

Mohan said query length has increased three times, and Marvin repeated it in the wrap-up. Google's own AI Mode report from May 2026 frames it narrowly — queries in AI Mode are on average three times longer than traditional searches — which is the version to quote, because it is an AI Mode comparison, not a claim about all of Google.

The queries are net-new, not refined

The detail with the biggest strategic consequence is one Mohan said she had not explained publicly before. Users are, in her words, "asking net new types of queries that they would have never thought to ask before." When Marvin asked whether follow-ups happen because the first answer failed, Mohan said no — and explained how Google measures it: "In our data science practice, we penalize queries, which we call man refs, manual refinements."

Two implications follow. First, Google optimizes against queries that need a second try, so the reward goes to answers that resolve intent in one shot — a coverage problem before it is a bidding problem. Second, conversational growth is not people rephrasing the same keyword three ways; it is demand that never appeared in your keyword tool because nobody typed it into a search box a year ago.

They are upper-funnel, and Google knows it

Mohan described queries "moving up the funnel," with more brainstorming questions and more questions about how to think about a category altogether. Schindler's framing matched it precisely: "On AI Mode, many of these queries are upper-funnel queries, allowing us to capture new high-value commercial intent," adding that "the majority of AI Mode queries are incremental."

That is the strategic gift of the fortnight. Google has said, in plain language, that the incremental query supply is research-stage and question-shaped. If your keyword plan is a list of product-plus-modifier terms, you are bidding on the shrinking half of the market.

Why This Breaks Classic Keyword Research

Three mechanics make old habits actively counterproductive now.

The eligibility trap: exact match cannot buy the new placement

Ads inside AI Overviews and AI Mode are matched to both the query and the content of the AI response. According to Google's documentation, only broad match keywords or keywordless targeting — AI Max for Search, Performance Max, Shopping and Dynamic Search Ads — can serve inside the overview. Exact match can still appear above or below it, and an exact match keyword does not block a broad match version from triggering inside.

Read that as a media plan. A non-brand campaign built entirely on exact and phrase match — the structure most performance teams were taught to build — is structurally opted out of the placement Google is scaling into ten new countries this quarter. Your AI-surface performance may be zero not because it failed, but because you never entered the auction.

Research prompts dominate the inventory that exists

Adthena's Holiday AI Paid Search Report counted approximately 175,100 paid ad placements inside Google's AI Overview and AI Mode panels in the US between June and September 2026: apparel about 111,000, consumer electronics about 55,000, travel around 6,500, beauty roughly 2,100, and toys and games around 685.

The same report classified the prompts those ads ran against. In beauty, 83% of ads ran on research and discovery queries — people comparing or asking what is best rather than naming a product. Travel was at 69%, toys and games at 67%, while apparel (48%) and consumer electronics (42%) sat just under half.

If four out of five ad impressions in a category attach to "what is best for X" rather than "buy X", the asset worth building is not a list of product names. It is an organized map of comparison and question language.

Your search terms report no longer separates the surfaces

Google includes searches that trigger ads in AI Overviews and AI Mode in aggregate within standard reports for all of Search, so AI-surface queries are intermingled through your search terms report. Combine that with mostly long, conversational incremental volume, and a blended CPA becomes close to meaningless: a 25-word research query and a three-word transactional query sit in one spreadsheet, judged by one number.

The fix is mechanical. Export search terms, segment by word count — one to three, four to seven, eight or more — then compare conversion rate, cost per acquisition or ROAS, and how often each segment appears in conversion paths. Long queries that look weak on last click but show up early in paths are your upper-funnel evidence; queries that appear nowhere are discovery spend you cannot yet justify.

A Five-Step Playbook for Rebuilding Keyword Research

Step 1: Rebuild the seed list around question chains

Stop seeding with head terms and modifiers. Seed with chains: what it is, which is best, X versus Y, is X worth it, what fits my situation. For each chain, capture the follow-up question a real person would ask next, because with multi-turn ads every step is its own ad slot — and organically, every step is a page you either own or forfeit.

Step 2: Group by decision stage instead of search volume

Volume sorting puts you back at the bottom of the funnel. Group the same set by stage — orientation, comparison, validation, purchase — and you will find both content and campaigns over-weighted at the end. The Adthena prompt data says the inventory is at the front.

Step 3: Give broad match or AI Max a defined slice, with clean data behind it

Since only broad match and keywordless targeting can serve inside AI answers, carve out a deliberate share of non-brand budget for them, backed by Smart Bidding and conversion tracking you trust. Keep exact match where you need control. The failure mode is not broad match itself; it is broad match with muddy conversion data and no negative keyword discipline.

Step 4: Feed the machine the attributes it now reads

For ecommerce, Google rolled out six optional conversational attributes in Merchant Center at Google Marketing Live in May: question_and_answer, document_link, related_product, item_group_title, variant_option and popularity_rank. They do not affect product approval, and Google recommends adding them through a supplemental data source so the primary feed stays intact. Mohan's note that comparison shoppers "really don't want to just see the summary" — they want reviews and videos from people in their own circumstances — is effectively what those fields are for. For content teams the equivalent is unglamorous: put the comparison table, the sizing detail and the use-case language on the page.

Step 5: Instrument AI surfaces as their own test line

Until Google reports AI Mode ad performance with the same durability standard it applies to user statistics, treat AI Mode and AI Overview reach as a test line with its own budget and its own success metric rather than something that quietly drags down blended CPA. Watch long-query share of clicks, assisted conversions, and cost per incremental conversion — not last-click ROAS.

Use Social Search as Your Prompt Lab

Here is the practical problem with building a question-chain set: the good questions are not in keyword tools yet, because they are new and conversationally phrased. Keyword databases lag behavior by months; social platforms do not.

Social search is the cheapest live corpus of research-stage language available. Comment threads, video questions and community posts are where people write the long, circumstantial phrasing — "I have a small kitchen and two dogs, is a robot vacuum worth it" — that never gets typed into a keyword planner but increasingly gets typed into an AI box. Mining that language gives you the prompt shapes to target, the objections to answer on-page, and the comparison axes your feed attributes should cover.

That is what SocialKeywordGenerator is built for: pull keyword and hashtag language from social platforms, cluster it by intent and question type, then hand those clusters to content and paid teams as prompt-ready sets rather than flat keyword lists. Start with SocialKeywordGenerator and build the question-chain map before your next planning cycle, not after it.

The Numbers That Keep You Honest

The up-funnel story is real, but it is not a licence to rewrite everything this quarter. The measurement data from the same fortnight is blunt about scale.

Marie Haynes' analysis of 12 months of GA4 data across 100 websites and more than 145 million sessions found ChatGPT sent 0.62% of the traffic volume Google organic sent, and all LLMs combined sent 0.73% — organic still delivered 39 times more traffic. ChatGPT visitors converted at 13.8% against 3.02% for Google organic, the stat vendors quote, but Google still drove 96.5% of total conversions, and Google organic drove 27 times more conversions than all LLMs combined.

So the correct read is not "AI is the new channel". It is "the query mix is shifting before the click volume is", which is precisely when keyword research decisions get locked in for the next year. Build the question-chain map now, while the inventory is cheap and the competition is still buying three-word terms. The surfaces are growing fast: Adthena counted US AI Overview appearances rising from 304 million to 575 million between August 2025 and August 2026, an 89% increase, with the UK and EU up 160% and Australia up 536%.

Key Takeaways

Google now monetizes AI Mode with multi-turn ads, ten more countries, home services booking links and product-feed creative — the commercial layer sits on conversational, research-stage queries.
Google's own data science says AI Mode queries run three times longer, are often net new, and skew upper-funnel; Google even penalizes "man refs", queries that needed manual refinement.
Only broad match or keywordless targeting can serve ads inside AI Overviews, so exact-match-only campaigns are excluded from the placement Google is scaling.
Research and discovery prompts carried 83% of beauty AI ads, 69% of travel and 67% of toys, versus 48% in apparel and 42% in consumer electronics.
Segment search terms by word count to see upper-funnel contribution; a blended CPA now hides both the surface and the stage.
ChatGPT still sends under 1% of Google organic's traffic volume, so fund AI surfaces as a measured test line, not a replacement.

FAQ

Q: What are multi-turn ads in Google AI Mode?

A: Ads that can appear on follow-up queries inside the same AI Mode conversation, not only on the first query. Google began rolling them out in the US on October 28, 2026, which means each step of a question chain becomes its own ad slot.

Q: Can exact match keywords serve inside AI Overviews?

A: No. Only broad match or keywordless targeting — AI Max for Search, Performance Max, Shopping and Dynamic Search Ads — can serve inside AI Overviews and AI Mode. Exact match can still appear above or below the AI answer, and does not block a broad match keyword from triggering inside it.

Q: Do I need to abandon bottom-funnel keywords?

A: No, but stop treating them as the whole plan. The incremental query supply is research-stage, so a plan without a question-chain layer competes for the smaller half of demand.

Q: How do I measure AI surfaces if Google blends them into search terms reports?

A: Segment by query length: one to three, four to seven, and eight or more words, then compare conversion rate, CPA or ROAS and appearance in conversion paths. Pair that with assisted-conversion reporting so upper-funnel touches stay visible instead of being flattened into a blended number.

Q: Which markets have AI Mode ads now?

A: Ads in AI Mode scaled to ten additional countries in late October 2026, including Canada and Australia, with Brazil, Mexico and the United Kingdom planned for Q4 2026.

Conclusion

The uncomfortable part of this fortnight is not that Google changed an ad format. It is that Google described, unusually directly, where its incremental queries come from — long, conversational, research-stage, often net-new — and then built ads to sit inside that conversation in ten more countries. Keyword research built on short commercial terms is now aimed at the part of the market that is not growing.

The response is not a bigger budget on every AI surface. It is to change what counts as a keyword asset: question chains instead of head terms, decision stages instead of volume tiers, prompt language mined from social search instead of only legacy databases, and reporting that separates long queries from short ones so upper-funnel work can be defended with evidence.

Start with the cheapest move: for your top ten categories, write out the five questions a real buyer asks before they type a product name, then check whether you own a page and a keyword for each step. In most accounts, two of the five are missing. Build that map with SocialKeywordGenerator, and you will be buying the inventory Google is scaling rather than bidding against the inventory it is leaving behind.