AI Mode users have stopped typing keywords and started naming exact attributes. Instead of "flights to Lisbon", they ask for "nonstop flights to Lisbon under a set budget in March"; instead of "restaurant near me", they ask for one "with outdoor seating and vegan options". This is a direct consequence of how Google AI Mode changed search behaviour, and it changes which pages get cited: the ones that expose the exact attributes a model can match against.
This piece explains the shift and what to do about it. It draws on Google's own AI Mode usage data (dated mid-2026, platform-reported, not independently audited) and applies the corpus method, so read the direction as firmer than any single number. For the raw figures see the AI Mode usage data reference; this is the "why it matters" companion.
Why did search shift from keywords to attributes?
Because the query format changed. A keyword search compresses intent into two or three words and leaves the rest implied; a conversational AI Mode query states the whole requirement in a sentence. Google reports three behavioural shifts that push people toward naming attributes:
| Shift | Google's figure (mid-2026) | Why it produces attributes |
|---|---|---|
| Longer queries | Average AI Mode query is about 3× the length of a classic search | More words means room to name filters, not just a topic |
| Conversational follow-ups | Follow-ups grew more than 40% per month | Each turn adds or refines an attribute |
| Multimodal input | More than 1 in 6 searches use voice or images | Spoken and image queries carry natural, detailed phrasing |
Source: Google, AI Mode U.S. Insights report (mid-2026). The pattern is the same across the query fan-out: one request splits into attribute-level sub-questions ("nonstop?", "under budget?", "in March?"), and the engine assembles the answer from pages that satisfy each attribute. This is the search-behaviour side of multimodal search: people say more, so they specify more.
What does an attribute query look like across verticals?
The same behaviour shows up wherever people use AI Mode to decide or plan. Google's report and trade coverage of it name the concrete attributes people specify in three categories:
| Vertical | Attributes people specify | Source |
|---|---|---|
| Shopping | Price, location, colour, brand, availability, size, material, style, type, quality | Google, AI Mode U.S. Insights (mid-2026) |
| Dining | "Outdoor seating", "vegan options", "private party room" | Google report, as summarised by CMSWire (mid-2026) |
| Travel | Nonstop, budget band, season, family-friendly, walkable, accessible | Illustrative, from Google's travel/planning findings |
The shopping attributes are Google's own ordered list, reused from the usage-data reference; the dining examples come from trade coverage of the report, so treat them as directional; the travel examples are illustrative of the same pattern, not a Google-published list. The through-line: whatever your category, buyers now state the properties they filter on, and the page that names those properties is the one a model can match.
When a query names an attribute, only a page that names the same attribute can be lifted to answer it. Vague copy is not just weaker: it is invisible to the sub-question, because there is nothing for the model to match.
Why do vague pages lose attribute queries?
Because retrieval and citation happen at the passage level, and a model matches an attribute in the query against an attribute stated on the page. "Great for families" cannot answer "hotel with a pool and connecting rooms"; "premium materials" cannot answer "waterproof and under a set price". The specific loses to nothing here, because the attribute simply is not present to be found. Three failure modes are common:
- The attribute lives only in an image (a spec chart, a menu photo), where a text-first crawler cannot read it.
- The attribute is implied, not stated ("comfortable for longer trips" instead of a named seat pitch or legroom figure).
- The attribute is buried mid-paragraph instead of in a table or list where it can be extracted cleanly.
Exposing attributes is the same discipline as getting cited by AI: make the facts explicit, corroborated, and structured.
How do you expose the attributes AI engines match against?
State them as facts, in the formats a model can lift. This is the cross-vertical version of the product-attribute work in the shopping playbook and the place-attribute work in the travel playbook:
- List the attributes your buyers actually filter on. Price, size, availability, dietary options, season, accessibility, whatever the category's attribute queries name.
- Put them in text and tables, not only images. A spec table beats a spec graphic every time for extraction.
- Add structured data (Product, Offer, or the relevant type) so machines read the attribute as a labelled fact, not a guess.
- Answer the attribute follow-ups, since AI Mode users refine conversationally: cover "and under budget?", "and nonstop?", "and vegan?" as their own chunks.
- Keep volatile attributes current and dated, because live retrieval favours recently-reviewed pages.
How should you measure this?
Track whether you appear for the attribute-level questions, not just the head term. A brand can rank for "hotels in Lisbon" and vanish the moment the query adds "with a pool and family rooms", because the page never named those attributes. So measure recall across the fan-out of attribute follow-ups, and do it per engine and over time rather than spot-checking one prompt. That is exactly what Buffy Intel is built to measure: whether AI Mode and every other major engine surface, cite, and recommend your brand for the specific, attribute-rich questions your buyers actually ask.