Buyers used to compare options. Now they ask an AI which options are worth comparing, and they trust the answer. That answer is assembled in real time, mostly from sources the brand does not own, not from your website or your ad budget. The brands that win the next decade are the ones easiest for a machine to retrieve, verify, and explain.
Source: AIVO audit data, ~1M citations across 6 engines (June 2026); aiplusautomation (Lee, 2026); Ahrefs (May 2026)
Dan Muirhead · Co-Founder, Head of Strategy
At AIVO we spend our days inside the AI discovery layer, watching how AI engines surface, describe, and rank brands across roughly millions of citations from hundreds of thousand audit records. The pattern that shows up first, before any nuance, is this: when someone asks AI an unbranded question about a category, the AI answers by naming a short list of specific companies. It does this about 72% of the time, and it names about four of them. A consideration set gets built in the time it takes to read a paragraph, and it gets built whether or not you are in it.
Nate B Jones put a clean name on what this adds up to. In May 2026 he argued that the web is moving from an attention economy, where the job was to get seen and get clicks, to an interpretation economy, where the whole web is filtered through what an AI thinks about you. We had been looking at the data that proves the shift, but he gave the shift its frame. The version of the story most of the industry tells is that this is an SEO problem with a new acronym. That version contains something real. Search and Digital fundamentals still matter, and a brand’s own pages still earn citations.
The complication is that the interpretation is mostly built from things the brand does not control, and it is assembled at the moment of the question rather than absorbed from your marketing in advance. The AI does not read your campaign and form an opinion. It reads whatever it can retrieve and trust about you across the open web, much of it third-party, and it composes an answer on the spot. This is why the most common instinct, publish more pages and mark them up with schema, underperforms. It treats the problem as one of supply when the problem is one of corroboration.
What follows is the research on how the interpretation economy actually works: how the consideration set forms, why the SEO-shaped response falls short, where the AI’s read of a brand really comes from, and what marketing leaders and operators across hotels, cruise, ecommerce, retail, SaaS, and beyond should do about it before the habit hardens. The single most expensive misconception in the market right now is that you get into the answer by getting into the model. You do not. You get in by being the easiest brand for a machine to retrieve, verify, and explain.
For twenty-five years the web ran on attention. Search was only the entry point. Once a buyer started looking, a category threw everything it had at them: paid ads, ranked search results, review sites, YouTube explainers, editorial roundups, comparison blogs, influencer posts, retargeting. The job of marketing was to win a share of that attention and convert it. The job of the buyer was to wade through all of it with finite time and patience, reading, watching, and comparing until they had assembled a shortlist of their own. Whether the purchase was a pair of running shoes or a six-figure software contract, the mechanics were the same: get seen across enough surfaces, get chosen. Attention was the scarce resource, and the entire commercial apparatus of the internet was built to capture it.
That work is being delegated. The buyer no longer wades through the ads, the roundups, the reviews, and the videos themselves. They ask an AI, and the AI does the wading. It reads the need, consumes the same sprawl of content the buyer used to consume, evaluates the field, and returns a short answer that names a handful of companies worth considering. The human shows up at the end of the process rather than the beginning. In AIVO’s audit data across 105,571 unbranded queries, the AI names at least one specific brand 71.9% of the time, and when it names brands it names about four. The consideration set is no longer something the buyer assembles by browsing. It is something the AI hands them, already narrowed.
| Query type | Answers naming a brand | Avg brands named |
|---|---|---|
| Branded (you asked for a company) | 99.3% | 3.7 |
| Unbranded (you asked about the category) | 71.9% | 4.2 |
This is the structural inversion. Attention is still finite, but a layer now sits between the buyer and the field of options, and that layer decides who is worth surfacing before the buyer evaluates anything. A brand can run a flawless attention strategy, win the impression, win the click in traditional search, and still never enter the set the AI builds, because the AI is not consulting your ad spend. It is consulting its read of you. The scarce resource is no longer the buyer’s attention. It is a seat in the four-name list the machine produces, and that seat is awarded by a process most brands are not yet managing.
“The brand is no longer what you say it is. It is what the AI says it is, assembled from what other people have published about you, at the moment of the question.”
Faced with this shift, most of the market has reached for the playbook it already owns. Treat it like search. Publish more content, target more query variations, and above all add structured data, because surely telling the machine exactly what you are in machine-readable form is how you get read. The logic is intuitive. It is also, on the most-cited single tactic, wrong.
In May 2026, Ahrefs ran the controlled test the industry had been assuming the answer to. They tracked 1,885 pages that added JSON-LD schema markup between August 2025 and March 2026, matched each against control pages with similar citation histories that never added it, and measured what happened to AI citations across Google AI Overviews, AI Mode, and ChatGPT. The result was nothing. AI Overviews moved minus 4.6 percent relative to controls, a small decline. AI Mode moved plus 2.4 percent and ChatGPT plus 2.2 percent, both statistically indistinguishable from zero. Four separate tests told the same story.
“Adding schema produced no major uplift in citations on any platform.”
| Platform | Change in citations | Verdict |
|---|---|---|
| Google AI Overviews | -4.6% | Small but real decline vs controls |
| Google AI Mode | +2.4% | Indistinguishable from zero |
| ChatGPT | +2.2% | Indistinguishable from zero |
The reason is mechanical. When a searchVIU experiment fed pages to five AI systems in real time, every one of them extracted only the visible HTML and ignored the JSON-LD, the Microdata, and the RDFa entirely. The markup most brands are paying to add is invisible to the systems they are adding it for. Pages with schema do get cited more often, but that is a quality correlation, not a cause: authoritative sites tend to do everything well, schema included. The schema is a symptom of a good site, not the lever that makes a site cited.
“Schema markup is a byproduct of quality, not a driver of citations.”
The deeper problem with the SEO framing is that it answers a supply question when the real question is corroboration. Publishing more owned pages adds to what you say about yourself. The interpretation economy runs on what can be retrieved and verified about you from independent sources. Volume without provability gets flattened into the category average, because the machine has no reason to treat your unsupported claim as more credible than a competitor’s. The work that moves the answer is not more output. It is evidence the machine can extract and a source it can trust.
If the consideration set forms from what an AI can retrieve and trust, the next question is where that material comes from. Part of it is the brand’s own, and that part deserves its due. Across roughly a million citations in AIVO’s data, brands are well represented, but mostly off their primary website. Their YouTube channels, their LinkedIn and Medium articles, their claimed directory and marketplace listings, their product and documentation subdomains all show up, and on some engines they show up heavily. YouTube, LinkedIn, and Medium sit among the most-cited domains in the entire dataset, ahead of most news outlets. Brand-authored content is a real and sometimes large input. It simply arrives through the side doors more than the front one.
What the brand does not author is the layer that decides trust. Alongside that owned material sits a wider field the brand has no editorial control over: competitors’ domains, third-party reviews, community discussion on Reddit and Quora, editorial coverage, reference sites like Wikipedia, and a long tail of independent pages. On unbranded category questions, this is where the consideration set is actually adjudicated. Competitor domains alone are cited nearly twice as often on unbranded queries as on branded ones, because asking about the category invites the whole field in. The brand contributes to its interpretation. It does not get to write it.
A note on rigor, because the honest version of this is more useful than the dramatic one. Cleanly separating brand-authored content from third-party content is genuinely hard for anyone measuring this, because so much of it lives on shared domains: a brand’s own product video and a critic’s takedown are both youtube.com, a brand’s claimed profile and a customer’s one-star review sit on the same review site. Counted strictly by the domain a company owns and operates, owned content looks tiny. Counted by everything the brand actually authored across every surface, it is a meaningful share, and external studies bear this out: roughly 47 percent of Google AI Overview sources are brand-controlled, and Claude draws about 53 percent of its citations from owned material. The picture also shifts with the question. On branded queries, where someone already named the company, AI leans harder on the brand’s own pages and on review sites, with review citations running about seven times higher than on unbranded queries, the reputation-defense layer at work. On unbranded category queries, where the consideration set is still forming, third-party corroboration dominates. So the claim is not that your website does not matter. It does. The claim is narrower and harder to escape: the mix that decides whether you enter an unbranded consideration set is engine-dependent, leans heavily on sources you do not control, and is never something owned content alone can win.
| Engine | Brand-controlled / owned | Earned & third-party lean |
|---|---|---|
| ChatGPT | lighter | ~51% earned/news; Wikipedia ~48% of top sources |
| Google AI Overviews | ~47% | YouTube ~30%; community and editorial fill the rest |
| Claude | ~53% | ~43% earned; rewards depth and structure |
| Perplexity | first-party + trade | recency-driven; community-led in external data |
| Gemini | mixed | trade, video, reference, structured comparison |
This reframes the work. Owned content still earns its place, and producing it well across your own surfaces is table stakes. But the leverage on unbranded discovery is not in adding more of your own voice to the pile. It is in shaping the sources the machine reads about you: the reviews, the editorial coverage, the community threads, the reference entries, the independent comparisons. There is a measurable threshold here. External analysis finds that only about 15 percent of the pages an AI retrieves actually get cited, and that presence across four or more independent authoritative surfaces is roughly the point at which a brand becomes citation-eligible. One owned page, however perfect, is one surface. The brands that come through are corroborated across many.
And there is a clear hierarchy in what earns the citation. Original research and proprietary data get cited at 38 to 65 percent rates in external testing, against 6 to 15 percent for blog content and 3 to 8 percent for product and marketing pages. The machine is not rewarding the brand that publishes the most. It is rewarding the brand whose claims are specific, dated, provable, and echoed by sources it did not have to take on faith. The interpretation is earned, in the literal sense.
The most expensive misconception in the market is that the way into an AI answer is to get into the model: feed it your data, publish enough that it learns you, train your way to relevance. The instinct is not baseless. There is brand information baked into a model’s training weights, and some engines lean on it more than others. ChatGPT and Claude run more parametric, answering a meaningful share of questions from what they already know, while Perplexity and Google’s AI surfaces lean on live retrieval. But even the parametric engines reach for the web when the stakes are high, and across the queries that build consideration sets, almost all of them go and read the open web at the moment of the question. A retrieval-augmented system reads sources in real time and composes an answer from what it just found and what it can verify. Your owned content is one candidate input among many, evaluated when the question is asked, not a lesson the model absorbed in advance and recites on command.
Intent is the other dial, and it is why a single tactic cannot work everywhere. The parametric engines still retrieve, but selectively. In a 391-query test of brand and product questions, ChatGPT triggered a web search only 42 percent of the time, answering the rest from training memory. The deciding factor was what the buyer was trying to do. Discovery questions, where a buyer is actively choosing, triggered search 73 percent of the time. Pure informational questions triggered it 10 percent of the time. The queries that build consideration sets are exactly the ones that force the machine to go read the open web about you.
| Query intent | ChatGPT triggers a web search |
|---|---|
| Discovery ("best CRM for small business") | 73% |
| Review-seeking ("best service desk on Reddit") | 58% |
| Validation ("is Mailchimp good for beginners") | 44% |
| Comparison ("Salesforce vs HubSpot") | 29% |
| Informational ("what is a CRM") | 10% |
The engines also differ in how much they retrieve and how readily they name brands at all. In AIVO’s data, Google AI Overviews cite roughly 15 sources per answer while ChatGPT cites about 3, which means source and citation levers have far more surface area on retrieval-heavy engines and far less on the parametric ones. Claude is the clearest case: on unbranded queries it names the client brand 5.2 percent of the time, against roughly 11 percent for the other engines, because it leans on training memory and surfaces specific brands less readily. A brand that wants into Claude’s answers competes on entity authority and corroborated reputation, not on how many pages it shipped last quarter.
“Agents need you to prove it.”
That is the correction the “train your way in” crowd is missing. You cannot push your interpretation into the model. You can only make yourself the brand that is easiest to retrieve, verify, and explain at the moment a buyer asks, so that when the machine goes looking, the evidence is there, it is consistent across surfaces, and it is provable enough to trust. The work is structural. It lives in the sources, the entity, and the proof, not in the prose.
The interpretation economy rewards a different discipline than the attention economy did. The shift, for operators across hotels, cruise, ecommerce, and SaaS, is to stop optimizing for rank and start optimizing for retrieval, citation, and provability. Three calls follow directly from the data.
Map your interpretation sources before you publish another page. Run your priority unbranded category queries through ChatGPT, Perplexity, Google AI surfaces, Claude, and Gemini, and record what the AI names, which brands it names alongside you, and which sources it cites to do it. That set of sources, the reviews, the editorial outlets, the community threads, the reference entries, is your real interpretation surface. It tells you where the answer is actually being assembled, which is almost never your own website. You cannot manage what you have not mapped, and most brands have never looked.
Earn third-party validation across four or more independent surfaces. Citation eligibility, in the external research, turns on presence across multiple independent authoritative sources, not on the quality of any single page. This makes earned media, review presence, and community standing direct inputs to AI visibility rather than a separate brand budget. The brand cited in a Forbes feature, discussed on Reddit, reviewed on the category’s leading platform, and described consistently on Wikipedia is the brand the machine trusts. Concentrate effort on the specific surfaces your mapping showed the AI actually reads for your category.
Make every claim provable, and lead with the proof. Original research and proprietary data are cited several times more often than marketing prose, and answer-first content with dated, specific, extractable claims wins the citation the machine is willing to attribute. Replace “world-class service” and “industry-leading platform” with the number, the method, the date, and the verifiable detail. This is the work that survives both retrieval and compression, and it is the opposite of adding more schema to more pages.
| Call | Action | Deliverable |
|---|---|---|
| 01 | Map your interpretation sources | Per-engine record of what AI names and which sources it cites for your priority unbranded queries |
| 02 | Earn validation across 4+ independent surfaces | Earned-media and review plan targeting the specific sources the AI reads for your category |
| 03 | Make every claim provable, proof first | Rewrite of priority claims into dated, specific, extractable evidence; original data where you have it |
This is where AIVO does its work, mapping the interpretation, measuring it across engines, and building the provable presence that earns the citation. The point of the three calls is not the framework. It is the reorientation: in the interpretation economy, the job is to be legible and provable to a machine that is reading about you from everywhere except your own front door.
For twenty-five years the web rewarded the brands that captured the most attention. That era is closing. Buyers are handing the consideration-set step to an AI and accepting its read of who is worth comparing, and the AI builds that read in real time, mostly from sources the brand does not own. The shortlist forms in a paragraph. It names about four companies. It forms whether or not you are one of them.
Nate B Jones named the shift; the data shows its mechanics. The interpretation is not absorbed from your marketing, so you cannot train your way into it. It is not won by volume, so publishing more pages does not move it. It is not a markup problem, so schema alone does nothing. It is assembled from what can be retrieved and verified about you across many independent surfaces, which means the brands that win are the ones easiest for a machine to retrieve, verify, and explain. That is a different kind of work than the attention economy asked for, and it favors the operators who start now.
The attention economy rewarded the loudest voice. The interpretation economy rewards the most provable one. The brand that cannot be verified will be averaged into its category, and the brand that can will own the answer.
AIVO maps your visibility across ChatGPT, Google AI, Perplexity, and more — then builds the strategy to close the gaps.