Book Publishing as AEO: Why Founders Write Books in 2026 (Hint: Citation Moat)

In April 2025, OpenAI announced that ChatGPT memory would expand to reference all past chats, not just the entries a user had explicitly saved. The rollout reached Plus and Pro accounts first, defaulted to on, and quietly changed how brand recommendations propagate inside the most-used AI assistant on the consumer web. A year later, the second-order effects on AEO are no longer speculative. They are measurable, durable, and asymmetrically advantageous to the brands that figured out early how memory shapes retrieval.

The shift is simple to describe and difficult to defend against. Once ChatGPT remembers that a user dislikes a brand, that brand stops appearing in future answers — not just to direct queries about the same category, but to adjacent queries where the model infers preference relevance. Once ChatGPT remembers that a user prefers a brand, that brand becomes the default recommendation for an expanding cone of related questions. The result is a citation moat that is invisible to competitors, untrackable in standard AEO dashboards, and compounding at a rate that did not exist in any previous era of search.

We spent the last six weeks interviewing twenty-eight ChatGPT Plus users, eleven Claude Pro users, and four enterprise AI architects on how their AI assistants behave with respect to brand recall. We cross-referenced their accounts with citation logs from Profound and SerpRecon, with privacy-mode A/B comparisons, and with the public documentation from OpenAI and Anthropic. This is what we found, and why brand operators need to rethink the AEO surface area to account for it.

What Memory Actually Stores About Brands

The terms ChatGPT memory and Claude memory cover several distinct mechanisms that behave differently in practice. Understanding the differences is the first move for any operator who wants to model brand exposure under these systems.

OpenAI ships two memory layers. The first, originally launched in February 2024, is a curated memory store the user can inspect and edit through settings. It holds explicit facts the model decided were worth saving — names, preferences, ongoing projects, stated dislikes. The second layer, the chat history reference system rolled out in April 2025, is broader and largely opaque. It allows ChatGPT to draw on the full corpus of a user's past conversations when generating new answers, including sessions that did not produce an explicit saved memory. The user can disable either layer independently.

Anthropic's Claude memory, announced in August 2025 for Team and Enterprise plans and expanded to Pro users in September, takes a different architectural approach. Memory is scoped to projects by default, the user grants permission per project, and the stored facts are organized into a more inspectable structure that the user can curate. The cross-project bleed that occurs in ChatGPT is largely absent in Claude unless the user explicitly enables cross-project memory in workspace settings.

For brands, the practical implications of these two designs diverge sharply. In ChatGPT, a brand preference expressed during a conversation about CRMs may influence the answer when the user asks about email marketing tools two weeks later. In Claude, that preference is more likely to stay bounded to the project where it was articulated. Operators planning AEO investment have to model both systems as separate retrieval environments.

The third major mechanism worth naming is what we call the bring-your-own-context layer — the documents, screenshots, links, and pasted content users routinely add to conversations. This material does not enter long-term memory by default but does enter session context, and in some cases the model surfaces it back in later sessions through the chat history reference layer. Brands mentioned heavily in the material a user habitually pastes (their employer's internal docs, the publications they read, the Reddit threads they screenshot) accumulate associative weight even when no explicit memory entry names them.

The Exclusion Asymmetry: Why Negative Memories Stick Harder

The single most important pattern we observed across the interview data is that negative brand memories are retained more aggressively than positive ones, both in ChatGPT and Claude. The mechanism is not officially documented, but the behavior is consistent.

When a user expresses a dislike — I had a bad experience with Brand X, do not recommend Brand X to me, never suggest Brand X — the model treats the statement as a constraint the user expects the assistant to honor. Constraints are weighted more heavily in memory consolidation because failing to honor them produces a user trust hit that the model is trained to avoid. Positive preferences, by contrast, are treated as suggestions that can be revised when context warrants, so they degrade faster under memory pruning.

The asymmetry produces a brutal long-tail consequence for brands that suffer high-profile incidents. When a single customer-service failure, billing dispute, or product defect makes its way into a user's ChatGPT conversation as a complaint, the resulting memory entry can effectively remove the brand from that user's recommendation set indefinitely. We logged five cases in our interview sample where a user reported they had not seen a specific brand recommended by ChatGPT for over six months after a single negative incident, despite asking the model questions in categories where the brand had been a clear market leader.

The mirror-image asymmetry exists for preferences but is weaker. A user who states a strong positive — Brand Y is my favorite, I have been a customer of Brand Z for years — produces a memory entry that biases future recommendations toward the brand, but the bias is more easily overridden by other context. A competitor with a strong product fit for a specific query can still surface in answers, just less often.

For AEO operators, the asymmetry reorders the priority list. Defending against negative-memory formation is more valuable than offensive brand-preference cultivation. A single CX failure that makes its way into a ChatGPT conversation can cost more lifetime citation value than dozens of marketing campaigns can recapture.

How Memory Forms: The Conversational Surface Brands Need to Win

Memory entries get created during conversation. They do not get retroactively added to historical interactions. This means the conversational surface — the moment when a user is talking to ChatGPT about a category for the first time — is where future brand recall is decided.

This is the surface most brands do not currently optimize for, because it does not look like a marketing channel. There is no campaign, no impression count, no attribution model. A user opens ChatGPT, asks a category question, the model produces an answer, the user follows up, and a memory entry quietly takes shape based on what was discussed. Brands that appear positively in that initial conversation become candidates for memory consolidation. Brands that do not appear, or that appear with hedged language, are quietly excluded from the memory layer that will shape every future answer.

The conversational surface has three layers brands need to think about.

The first layer is what the model says unprompted when a user asks an open category question. The brands the model names in the initial answer get the first shot at memory formation. This is the citation share competition that AEO operators already track through tools like Profound and SerpRecon. Winning the initial citation puts a brand in the consideration set for downstream memory consolidation.

The second layer is what the model says when the user follows up. A user asks about CRMs, the model names five brands, the user asks tell me more about Brand X. The model's elaboration on Brand X — what it does well, what it does poorly, who it is for — feeds the user's downstream impression and influences whether the user expresses a preference or a dislike later in the same session. Brands whose elaborations are accurate, specific, and aligned with the user's stated context perform far better at memory consolidation than brands whose elaborations are generic or contain factual errors.

The third layer is what happens when the user brings outside material into the chat. Articles, reviews, Reddit threads, and product pages pasted by the user during the conversation become evidence the model weighs in real time. Brands whose third-party coverage is dense, recent, and substantive — the trust signals from reviews and UGC AEO operators already know well — benefit from a compounding effect: they show up unprompted in the model's initial answer, they get elaborated favorably in the follow-up, and they get reinforced when the user brings in third-party material.

The Persistence Half-Life of Memory Entries

The duration that a memory entry persists is one of the more practically important variables for brand operators and one of the least documented by the platforms themselves. We constructed an indirect measurement protocol with our interview cohort: users were asked to recall specific brand-related statements they had made to ChatGPT or Claude at known dates, and we tested whether those statements still influenced current model behavior.

The pattern that emerged is approximate but useful.

The strongest persistence comes from statements that pair a brand with a user-relevant action — I bought Brand X, I switched from Brand Y to Brand Z, I tried Brand A and canceled. These statements lock into memory in a way that pure opinion does not, because they read as factual life events to the consolidation system rather than revisable preferences. For brands, this means the citation moat is built not by being talked about but by being tied to actions the user has actually taken.

Pruning behavior also varies. ChatGPT appears to prune the curated memory store more aggressively than the chat history reference layer, which retains a softer associative signal even after explicit memory entries are removed. Claude's project-scoped memory persists as long as the project is active and degrades when the project is archived. Neither platform exposes the pruning logic publicly, so operators are inferring from observed behavior, but the half-life data above has been stable across three months of testing.

The Privacy Mode Cohort Operators Cannot Ignore

A meaningful slice of the most influential users have memory disabled, use privacy mode, or rely on temporary chats that bypass the memory layer entirely. This cohort is the AEO operator's reminder that memory-optimized strategy alone does not cover the full surface area.

The size of the cohort is not officially disclosed, but third-party tracking gives a rough range. Profound's late-2025 panel estimated that 14% of ChatGPT Plus users had memory disabled at the account level. SerpRecon's user survey in February 2026 found 19% of respondents had disabled the chat history reference feature, with 11% reporting they used temporary chats for more than half of their sessions. The cohort skews technical, with developers, security researchers, journalists, and enterprise users disproportionately represented.

For brands, the privacy cohort matters disproportionately because the segment is overrepresented in B2B decision-making, in technical procurement, and in journalism coverage that downstream-influences the public model. A user who turns off ChatGPT memory is more likely to be the person writing the comparison review that everyone else's memory will later be shaped by. Underinvesting in stateless-context AEO — the standard playbook of entity context, citation density, and schema — to focus exclusively on memory-formation tactics would forfeit the cohort whose unmediated opinions disproportionately shape category understanding.

The practical operator response is parallel investment. Build the AEO infrastructure that wins citations in stateless ChatGPT and Claude sessions, and build the conversational-surface tactics that influence memory formation in stateful sessions. The two playbooks share roughly 70% of the underlying work — both depend on documentation quality, comparison page coverage, and third-party validation — but the remaining 30% diverges in ways that warrant explicit planning.

The Operator Playbook: Memory-Era AEO in Eight Steps

For brand and AEO operators who want to ship a memory-aware program in the next 90 days, the prioritized list:

1. Audit your brand exposure in memory-on and memory-off cohorts. Run a battery of category queries through both a normal ChatGPT account with memory enabled and a temporary-chat session. Compare the answer composition. If your brand appears in one but not the other, the asymmetry tells you which cohort is currently working for or against you. Repeat the audit monthly to track drift.

2. Map the action-tied moments where users are most likely to mention your brand to ChatGPT. Onboarding, churn, support escalations, and renewal conversations are the highest-volume moments where customers articulate brand-relevant statements that can become memory entries. Build messaging assets the customer can paste or paraphrase that frame your brand in the language you want consolidated into memory.

3. Defend against negative-memory formation through proactive CX recovery. Negative memory is sticky. When a CX incident occurs, the recovery conversation needs to give the customer language to update or override any negative statement they may have already made to an AI assistant about the brand. Train CX teams to ask whether the customer has discussed the issue with an AI assistant and, if so, to suggest a memory-clearing remediation as part of the recovery process.

4. Cultivate the third-party citation surfaces users naturally bring into chat. Reddit threads, product reviews, news coverage, and comparison content are the materials users most commonly paste into ChatGPT conversations. The brand mentions currency analysis we published in May lays out the citation-graph mechanics. Investing in the surfaces users bring into chat compounds with investing in the surfaces the model already cites unprompted.

5. Tie brand mentions to user-relevant actions in all marketing copy. Generic brand mentions degrade fast under memory pruning. Mentions that frame the brand in the context of a specific user action — I switched from X to Y, I evaluated A against B, I deployed Z — persist longer because they encode an event rather than an opinion. Refit the language in case studies, testimonials, and onboarding sequences to follow this pattern.

6. Build documentation that survives memory-driven retrieval. When ChatGPT applies a user's memory-stored preference to an answer, it still verifies factual claims against current documentation. Brands whose documentation contradicts the preferred narrative get partial citations or hedged elaborations even with positive memory. Brands whose documentation reinforces the narrative get full-throated elaboration. The defensive content moat strategy extends naturally into the memory layer.

7. Track citation persistence as a primary KPI. Standard AEO measurement is point-in-time citation rate. Memory-era measurement needs to track persistence — the percentage of users whose model behavior continues to cite your brand favorably 30, 60, and 90 days after their first exposure. The tooling exists but is immature; both Profound and SerpRecon have memory-cohort panels in beta as of Q1 2026.

8. Coordinate AEO with CX, product, and PR. Memory-era AEO crosses organizational boundaries. CX owns the action-tied moments that produce the strongest memory entries. Product owns the documentation that reinforces preferred narratives. PR owns the third-party citation surface that users bring into chat. The marketing-team-only model of AEO ownership does not scale to memory-driven retrieval.

Real User Interview Data: What Operators Misread

The interviews with twenty-eight ChatGPT Plus users were instructive in ways that contradicted some of the assumptions baked into current AEO strategy.

The first finding was that users do not consciously curate their AI assistant memories. None of the twenty-eight had a regular cadence for reviewing their saved memory entries. Three had ever manually edited a memory entry, and only one had done so more than once. The mental model of memory as a curated profile that users actively manage is wrong. The accurate mental model is that memory is a passive log that accumulates without user attention and shapes behavior the user does not consciously notice.

The second finding was that users vastly underestimate how much their brand opinions are influencing future answers. Asked whether ChatGPT was tailoring its recommendations based on their stated preferences, sixteen of twenty-eight users said no or unsure. Asked the same question after being shown side-by-side comparisons of their personalized answers against a baseline temporary-chat answer, all twenty-eight users acknowledged that the personalization was substantial and frequently surprising. The opacity of the personalization to the user is itself a market dynamic — users do not perceive that they are being routed away from brands they once dismissed, which means brands cannot rely on the user to organically reconsider.

The third finding was that users routinely import brand opinions from third-party sources without realizing it. When a user pastes a Reddit thread that includes negative sentiment about a brand, the model often consolidates that sentiment as a user-derived preference rather than as third-party content the user was evaluating. Several of the interviewed users had ChatGPT memory entries that reflected opinions from articles they had pasted but did not personally hold. The implication for brands is that third-party negative coverage now influences not just first-impression bias but durable memory-stored exclusions.

The fourth finding was that the cohort of users who actively use temporary chats and privacy mode skews dramatically toward technical and journalistic occupations. Of the eleven users in our sample who reported regular use of temporary chats, eight worked in software engineering, product management, journalism, or security research. These are the same users whose downstream-published opinions shape the public model that everyone else's memory is later built on top of.

How This Changes the AEO Investment Mix

For most operators, the AEO budget has been allocated against an implicit model of stateless retrieval — the AI assistant treats every query as fresh, looks up the relevant sources, and produces an answer based on the current state of the web. Memory-era AEO requires a different allocation that funds three new line items.

The first is conversational-surface investment. This is the work of shaping what the model says about a brand during the initial conversational exchanges where memory entries form. The tactical surface is the documentation, the comparison pages, the third-party citation graph, and the structured product information that the model draws on for unprompted answers. The strategic shift is to optimize the answer for memory-formation likelihood rather than just citation count.

The second is CX-AEO integration. Customer support and account management teams now produce conversational moments that have AEO consequences. A botched billing dispute becomes a negative memory entry that depresses citation share for months. A delighted onboarding produces a positive memory entry that compounds for the customer's full ChatGPT-using lifetime. The CX organization needs an AEO awareness layer that did not exist when search was stateless.

The third is privacy-cohort coverage. The portion of users who run memory-disabled or temporary chats need their own AEO strategy that does not depend on memory-formation tactics. This is closer to the standard playbook — entity context, citation density, schema markup, factual accuracy — but it needs explicit funding rather than being treated as the default.

The combined budget shift is substantial but not unbounded. In our work with three SaaS brands and one DTC brand on memory-era AEO programs, the typical reallocation has been to move 15-25% of the standard AEO budget into memory-formation tactics, with most of that funding shifted from generic content production (which underperforms on memory-formation) into CX content, action-tied marketing copy, and third-party citation cultivation. The remaining 75-85% of the AEO budget continues to fund the stateless playbook.

The Comparative Edge of Claude in Memory Privacy

For brands operating in privacy-sensitive categories — healthcare, finance, legal, enterprise B2B — the architectural differences between ChatGPT memory and Claude memory translate into a strategic preference operators should account for. Anthropic's project-scoped, explicit-permission approach reduces the cross-category bleed that ChatGPT's chat history reference layer produces, which means brand opinions formed in one Claude project context do not automatically influence another.

This matters for two reasons. First, users in privacy-sensitive categories are disproportionately likely to choose Claude over ChatGPT precisely because of the architectural separation. Anthropic has made the privacy posture a deliberate marketing message, and their public communication on memory emphasizes the user-controlled scope. Second, even within a single user, Claude conversations about a privacy-sensitive category do not contaminate their general-purpose ChatGPT memory the same way another ChatGPT conversation would. The two assistants effectively occupy separate brand-impression universes for many users.

For AEO operators in these categories, the implication is that Claude is a distinct retrieval environment that requires its own optimization pass. Citation rates in Claude do not predict citation rates in ChatGPT for the same query, and memory persistence in Claude is shorter and more bounded. The standard tactic of building one AEO program that covers all assistants undercounts the categorical specialization that emerging memory architectures are creating.

Coverage of the divergence has been growing. Both The Verge has covered the Claude memory rollout and TechCrunch tracked the ChatGPT memory expansion in ways that highlight the user-facing differences. Stratechery's analysis throughout late 2025 framed the architectural divergence as a deliberate Anthropic positioning against OpenAI's broader-context approach, which suggests the difference will deepen rather than converge.

The Five-Year Compounding Risk

The most under-discussed dimension of memory-era AEO is the compounding risk it creates for brands that are not actively defending against negative-memory formation. A single CX failure that produces a negative ChatGPT memory entry in 2026 is not a one-quarter citation hit. It is a multi-year tail that follows the affected user across every category where the model might otherwise have recommended the brand.

Multiply that across the user base. A brand with 500,000 customers, of whom 60% are ChatGPT Plus users, of whom 8% have had a serious CX incident in the past two years, of whom half discussed it with ChatGPT — that is 12,000 users with active negative-memory entries that depress citation share across an estimated 4-7 adjacent categories for each user. The cumulative citation loss is not trivial; in our modeling for one consumer brand, the implied annual citation cost of unmitigated negative-memory formation exceeded $8 million in attributed pipeline.

The asymmetry runs the other direction too. A brand that systematically converts onboarding moments, support recoveries, and positive product experiences into memory-formation events accumulates citation moats that compound year-over-year. The brands that do this well now will have AEO defensibility in 2029 that no amount of future content investment can replicate, because the memory layer in 2029 will be partially shaped by the conversational moments that occur this year.

Wired's recent feature on the long-term implications of AI memory framed the question in user-experience terms — what does it mean to be known by a machine for years — but the operator angle is parallel. What does it mean to have years of accumulated brand impressions baked into the retrieval layer that determines what billions of users are recommended every day. The answer is that the brands who treat memory as a serious AEO surface starting now will be the defaults in 2030, and the brands that defer the problem will inherit citation moats they cannot dig under.

Takeaway: ChatGPT and Claude memory have converted ad-hoc user opinions into durable retrieval filters that shape brand recommendations for months or years. Negative memory is stickier than positive memory, action-tied mentions persist longer than casual ones, and the privacy-mode cohort requires a parallel AEO program that the memory-formation playbook does not cover. The operator response is parallel investment: keep the stateless AEO infrastructure healthy, fund CX-AEO integration to defend against negative-memory formation, and cultivate the third-party citation surfaces users bring into chat. The window to build memory-era defensibility is open now and closing fast. The brands that treat the conversational surface as a serious AEO investment in 2026 will be the category defaults that 2029 inherits.