The Local Buyer's New Consideration Set: Second-Level Agenda-Setting in the Map Pack and the AI Answer

The consideration set, and who now assembles it

In consumer research a consideration set is the small subset of options a buyer actively weighs before choosing, not every option that exists. The set has always been small because attention is scarce. Herbert Simon named the economics of that scarcity in 1971: a wealth of information creates a poverty of attention, and a decision-maker must allocate attention among an overabundance of sources that would otherwise consume it.

Simon also gave us the reason the assembler of the set holds so much power. His account of bounded rationality holds that people satisfice rather than optimize: they act on the first answer that is good enough, not the demonstrably best one. A synthesized answer that names three credible businesses is, for most buyers, good enough. The step a buyer used to perform by scanning a page of results, narrowing many options to a few, is now performed upstream by the engine.

The set the engine returns is genuinely small. Pew Research Center's 2025 browsing-panel study found that when an AI summary appeared, 88 percent of those summaries cited three or more sources and only 1 percent cited a single source. A handful of names, assembled by the engine, is what most buyers now see and act on.

First-level agenda-setting: which businesses are named

The precise vocabulary for this comes from communication science, not marketing. In 1972 McCombs and Shaw showed that the media do not tell people what to think so much as what to think about: by selecting which issues to cover, they transfer salience to those issues and set the public agenda. The mechanism is selection, not persuasion.

That is the first level of agenda-setting, and it maps cleanly onto local search. The map pack names three businesses in a boxed result above the links. The AI answer names a handful inside a written response. Neither ranks the full field for the buyer to browse; each performs a selection and hands back a short list. Being absent from that list is not the same as ranking eleventh. It is being left out of the consideration set entirely, before price, reviews or years in the trade ever enter the buyer's mind.

Karine Barzilai-Nahon formalized the general case as network gatekeeping in 2008: the power to include, exclude and order the information others depend on, defined by who controls a gate and how much the gated depend on it. A local buyer depends heavily on the map pack and the answer, and controls neither. The engine is the gatekeeper of the consideration set.

Second-level agenda-setting: which attributes are emphasized

The move that matters most for a business trying to get chosen is the second one, and it is the least discussed. Alongside object salience, agenda-setting theory documents attribute salience, its second level: media influence extends beyond which objects an audience considers to which characteristics of those objects seem important. When coverage repeatedly frames a candidate by competence rather than warmth, competence becomes the trait the public weighs. Salience transfers at the level of attributes, not just objects.

The map pack and the AI answer do exactly this. They do not simply name a business; they frame it. A map result foregrounds a star rating, a review count, a category label and sometimes a response indicator. An AI answer describes each named business in a clause or two, and those clauses select attributes: highly rated, board-certified, family-owned, open now, budget-friendly. The buyer inherits those attributes as the criteria of choice. If the engine emphasizes reviews, the decision becomes about reviews. If it emphasizes credentials, the decision becomes about credentials.

This is why two businesses can both be named and still not compete on equal terms. The one whose foregrounded attributes match what the buyer was already primed to want wins the comparison the engine has quietly framed. Second-level agenda-setting predicts that the fight has two parts: being in the set, and being described in it by the attributes that decide the category.

Why high-consideration local verticals are where this bites

Attribute framing matters most where the buyer cannot easily verify quality before purchase. Med-spa treatments, emergency home services, dental work and legal representation are high-consideration, trust-dependent decisions. The buyer leans on proxies, and the engine supplies the proxies it chooses to foreground. In a med-spa answer the emphasized attribute might be safety and credentials; for a burst pipe it is availability and speed; for a lawyer it is track record and specialization. Whichever the engine repeats becomes the axis the buyer decides on.

The structural escalation is the collapse from a comparable list to a single framed answer. Eli Pariser's 2011 account of the filter bubble described personalization deciding what a person is even shown; a ranked list of ten still let a buyer compare. A synthesized answer shows one framing, with no ranked alternatives beside it to contest the attributes it chose. For a locally chosen business, that framing is often the whole decision, made inside a result the owner never sees rendered.

None of this requires the engine to be right. Simon's satisficing already told us the buyer will act on the first adequate, framed answer rather than audit the field. The business that shaped the attributes in that answer captures a disproportionate share of the decision, whether or not it is, on the merits, the best choice available.

The evidence that the gate has narrowed

The claim that the named set now decides more of the outcome is not a vibe; it shows up in primary and independently corroborated data.

What the theory proves for local, and what it does not yet

Agenda-setting is established: it is among the most replicated theories in communication science, with first-level, second-level and network refinements documented across five decades. That the map pack and the AI answer perform selection and framing is well supported by the theory and by the click data above.

The application to generative local answers is emerging, not settled. A 2026 peer-reviewed cross-language study by Kuai and colleagues frames AI chatbots as meta-gatekeepers that gatekeep the web's existing gatekeepers, deciding which already-curated sources surface at all. It is methodologically serious but a single recent study that needs replication, and it studied political information rather than commercial local queries.

The specific number a local owner most wants, how often engines name businesses in their vertical and which attributes they foreground, is not in the public record. No rigorous, methodology-disclosed count of local-intent AI citations broken out by med-spa, home services, dental or legal has been published that we can cite. This is a genuine gap. The theory tells us the mechanism exists; only direct measurement against a real panel of buyer questions can tell a given business where it actually stands. The response is to measure, not to assert a figure that does not yet exist.

The field's forecasting record is its own caution. Gartner predicted in 2024 that search-engine volume would drop 25 percent by 2026 as AI agents absorbed queries. As of this writing that has not materialized as stated, and Google still holds the large majority of search. The shift in who assembles the consideration set is real; confident totals and timelines are where predictions in this domain keep going wrong.

Engineering for the attributes that get named

If second-level agenda-setting is the mechanism, the practical question is which signals an engine reads before it frames a business. The academic literature offers one grounded answer. The 2024 peer-reviewed GEO study by Aggarwal and colleagues, published at KDD, tested content interventions for their effect on whether a source is cited inside a generated answer, and found that adding cited statistics, quotations and authoritative sources raised a source's visibility in the systems it tested. The object being optimized is no longer a rank; it is a citation, and the attributes attached to it.

Translated to a local business, that points to a few disciplined moves rather than any guarantee. Make the entity resolvable: one consistent name, address and category across the profile and the directories, so an engine is not hedging between three conflicting records. Corroborate the attributes you want foregrounded with third-party evidence, real reviews, verifiable credentials, named accreditations, rather than self-description an engine cannot trust. Structure the content so the claims are extractable. These are the inputs to attribute salience; none of them promises a ranking or a citation.

The through-line is that appearing and being framed are two different bodies of work. A business can rank on classic search, be absent from the AI answer, and lose the decision to a competitor the engine named and described more favorably. Working one surface while the others stay flat moves nothing the buyer sees.

What the evidence supports

The measured conclusion resists both denial and doom. Attention is being reallocated across more surfaces, classic search, the map pack, the AI answer and reputation, not vanishing from any one of them. The businesses that win are not the ones that panic about a collapse that has not arrived on schedule; they are the ones that show up, and are described well, across every gate a buyer now passes through.

That is also why a single ranking number is the wrong scorecard for this moment. Salience is transferring on multiple surfaces at once, and at two levels on each, which businesses are named and which attributes are emphasized. A standing worth trusting has to be read across all of them, on a real panel of the questions a buyer actually asks, and dated against the engines it was run on. The theory is settled enough to act on. The measurement is where the work begins.