In March 2025, Pew Research Center tracked 68,879 Google searches across 900 U.S. adults. When the search produced an AI Overview, users clicked a traditional result 8% of the time, against 15% for searches without one. They clicked a link inside the AI summary 1% of the time. The summaries themselves cited three or more sources in 88% of cases.
That is not a worse click-through curve. It is a phase change in how attention reaches brands.
For two decades, the search engine was an index that pointed users at pages where the work of persuasion happened. The page was the unit of competition. The answer is the new unit. The user receives a synthesized response, often without leaving the chat surface, and the brand either appears inside that response — or it doesn’t.
A second move is already underway, and it is the one that should keep brand teams awake. Tool-using language models, ReAct-style agents, computer-use APIs — autonomous systems that don’t stop at the answer. They book the flight, fill the cart, file the ticket, request the refund. The unit of competition becomes the executed action. The brand either gets selected, called, transacted with — or it doesn’t.
The construct that captures what is at stake here doesn’t yet exist in the marketing canon. I want to call it brand erasure, and the rest of this article is an argument for why the term deserves to be named, taught, and tracked.
A working definition of Brand Erasure
Brand erasure is the condition under which an AI assistant satisfies a user’s underlying need without surfacing, citing, or transacting with the brand whose content, product, or service the response depends on.
The definition is operational on purpose. It does not require the assistant to actively misrepresent the brand or to harbor any disposition toward it. Erasure is what happens when the brand is absent from the user’s experience even though the brand is present in the substrate the assistant draws on.
It is not brand dilution. Dilution describes what the consumer thinks of the brand. Erasure asks whether the consumer thinks of the brand at all.
Disintermediation removed the middleman. Erasure inserts a new one — an AI intermediary that can hide the brand from view as easily as it can connect to it.
Algorithmic invisibility describes content failing to surface in a results list. Erasure can happen after a brand surfaces: synthesis paraphrases the content into prose without attribution, or an agent picks a substitutable platform at the moment of action.
Three pathways, in increasing severity
Erasure operates through three mechanisms, each tied to a stage of the assistant’s procedure.
Extractive erasure. A user asks a closed-domain question. The assistant lifts a fact from the brand’s content without naming the source. The consumer gets the information; the brand provides the value and gets none of the attention. This is the smallest form, and the one most continuous with classical SEO disintermediation.
Synthetic erasure. A user asks an open-domain question. The assistant retrieves multiple sources and produces a synthesis. Brand content survives — paraphrased — but attribution does not. Liu, Zhang, and Liang’s 2023 audit of production generative search engines found that only 51.5% of generated sentences were fully supported by their citations, and only 74.5% of citations supported their associated sentence. The gap is not random. Paraphrase-friendly content gets recombined; paraphrase-resistant content keeps its citation.
Agentic erasure. The user delegates a task. The agent fulfills the category-level need by selecting a substitutable platform that may have no relation to the brand the user might have chosen if the user had been the active party. The brand is not delisted from a results page. It simply never enters the action.
Each pathway has its own defense, and getting the diagnosis wrong is expensive. AEO tactics — atomic propositions, schema markup — protect against extractive erasure but do nothing for the agentic kind. A platform with perfect agentic infrastructure can still be erased upstream, in the answer that decides whether anyone delegates to it in the first place.
Why classical brand theory runs out of road
Keller’s Customer-Based Brand Equity has been the most generative idea in brand theory of the last three decades. It defined brand knowledge as a network of associations in consumer memory, organized along awareness and image, varying in favorability, strength, and uniqueness.
Its foundational assumption is that brand outcomes are mediated through consumer cognition. The consumer encounters the brand, retrieves associations, evaluates options, chooses. The mediation chain runs through human memory.
In AI-mediated markets, an additional mediation layer exists. The assistant retrieves, evaluates, and chooses on the user’s behalf, drawing on a different substrate — the latent representations of foundation models, the entity graphs that ground retrieval, the structured manifests that machine surfaces expose. The consumer’s memory matters less if the assistant’s representation matters more.
The translation from brand-equity theory to the AI substrate is, at the level of structure, surprisingly direct. A brand association in consumer memory has a node, a connection, a weight. A brand association in the latent space of a language model has a vector position, a distance to attribute vectors, a magnitude. The mathematical structure is the same — a graph, plus weights, plus a similarity measure. What changes is whose graph it is.
When a user asks an assistant for “a reliable mid-range running shoe,” the relevant brands are surfaced via similarity to {reliable, mid-range, running, shoe} in the latent space. If your brand vector clusters with “discount” rather than “reliable,” your visibility loss happens before retrieval, in a layer the SEO industry has historically ignored. Brand-equity work and search-marketing work, long treated as adjacent crafts, become the same craft. The unit of analysis is the latent representation.
The optimization stack: AEO, GEO, AgO
Practitioners have responded with three acronyms: Answer Engine Optimization, Generative Engine Optimization, Agentic Optimization. The trade press treats them as separate playbooks. They are not. They are three layers of the same problem, each targeting a different stage of the assistant’s procedure.
AEO is the extraction layer. The target is closed-domain queries — questions with a single correct answer that can be lifted from a passage. The tactic is propositional density: short, declarative sentences that bind the fact to the brand’s name in the same proposition. The metric is fact-extraction accuracy.
GEO is the synthesis layer. The target is open-domain queries where the assistant builds a narrative from multiple sources. The unit of competition is citation worthiness. Two underappreciated levers drive it. Information gain — the marginal informational contribution your source makes compared to others retrieved for the same query. Definitional anchoring — what happens when your brand introduces or operationalizes terminology that becomes load-bearing in the answer. The metric is share-of-citation.
AgO is the action layer. The target is execution, not text. Tactics are infrastructural: published manifests, machine-readable schemas for product, pricing and inventory, idempotent endpoints, low-latency authentication, clear error semantics. The metric is agent execution rate — the share of tasks an agent attempting work on your surface successfully completes.
The strategic question is not which layer to invest in. It is which layer is currently rate-limiting for the brand’s category. Booking, ordering, and scheduling categories are AgO-bound. Considered purchases — B2B software, financial services, education — remain GEO-bound. AEO is universal but rarely decisive on its own.
The asymmetry that should worry challengers
A property worth establishing up front because it shapes everything that follows: brand erasure is asymmetric.
Established brands with high entity-graph presence — canonical Wikipedia entries, Wikidata identifiers, stable named-entity recognition across foundation models — are erased at substantially lower rates than challenger brands whose distinctiveness is carried in copywriting and visual identity rather than in machine-readable structure.
The legacy of brand-equity investment shows up in the AI substrate, but only if the investment was indexed. This inverts the conventional brand-management orthodoxy. Decades of brand work that produced a distinctive logo and memorable copy created equity in human memory. The same work might produce nothing in the latent space if the brand never showed up in the corpus, never anchored a category term, never built the entity-graph relationships that ground retrieval.
The implication for challengers is harsh: a louder voice in the click economy is not the same asset in the answer economy. The new asset is indexed presence.
The KPI shift
The KPIs of the click economy — sessions, click-through rate, bounce rate, time on page — are not wrong. They answer a question that is becoming peripheral.
Click-through rate gives way to share-of-citation: how often the brand appears in synthesized answers in its category. Page rank position gives way to embedding proximity: where the brand sits in the latent space relative to category attributes. Keyword density gives way to definitional anchoring: whether the brand’s terms become load-bearing in category answers. Conversion rate gives way to agent execution rate: whether agents successfully complete tasks on the brand’s surface. Brand search volume gives way to brand entity grounding: whether the brand has stable, canonical representation in public knowledge graphs.
None of these new KPIs is exotic. Each can be measured with audit methods adapted from existing toolkits — algorithmic-accountability audits, embedding-probe techniques from the bias literature, response-audit protocols. The reason they are not yet standard is organizational, not technical.
Two responses: anchoring and gatekeeping
Strategic responses to erasure divide along an open and closed axis.
The open strategy is semantic anchoring. The brand operates as if every public document is training data and every term it cares about is a candidate for definitional capture. It publishes its methodologies under named labels. It seeds its category vocabulary in academic, regulatory, and industry literature. It builds relationships in the entity graph through high-authority co-mentions. The point is to make the brand’s terms inseparable from the category’s terms, so that paraphrase cannot strip the brand from the synthesis without breaking the meaning.
The closed strategy is agentic gatekeeping. Premium data, proprietary insights, and competitive intelligence sit behind authentication. Agents that want access must identify themselves and accept attribution conditions. The risk of this strategy is also its mechanism: by forcing identification, the brand preserves its source identity at the cost of corpus presence. For categories where the asset is the proprietary insight — financial research, premium analytics, regulated data — this is a defensible trade. For commodity content categories, it is a way to disappear.
Most brands will need both, in proportions that vary by asset. Open content at the top of the funnel, where the goal is corpus presence and definitional anchoring. Closed content at the bottom, where the goal is forcing identification at the moment of value transfer.
The temporal window
The substrate AI assistants operate in is malleable now, more than it will be in five or ten years. Foundation-model training corpora are still being constituted. Entity graphs are still being built. The conventions of attribution that will harden into industry norms are still being negotiated.
Brand work that addresses this period has a chance to influence the conventions that emerge. Brand work that waits for the conventions to settle will arrive after the period in which it could have shaped them.
What the brand actually owns
The conclusion practitioners draw most often — build for agents instead of for people — is wrong in the same shape that “build for mobile instead of for desktop” was wrong a decade ago. The audience is plural by default. Every artifact you publish has at least two readers, and the machine reader sets the conditions under which the human reader gets to read at all.
What the brand owns, in the end, is not its place in the SERP, the answer, or the action. What it owns is its representation in the substrate beneath all three. Build the substrate, and the surface forms become a problem of distribution. Skip it, and the brand becomes the kind of thing the assistant can do without — which is the only outcome the framework was built to prevent.
Brand Erasure References
Figueira, M. G. (2026). Brand Erasure: A Research Agenda for Brand Theory in Agentic AI Markets. Fundação Getulio Vargas (FGV) & Wyse Brand Intelligence. Academia.edu
Figueira, M. G. (2026). From Information Retrieval to Agentic Action: A Framework for Brand Visibility in AI-Mediated Markets. Fundação Getulio Vargas (FGV) & Wyse Brand Intelligence. ResearchGate