Optimizing for People and Models at Once

The conversation inside most communications teams goes something like this: someone argues that content needs to be written for people, not machines. Someone else argues that if AI systems don't surface it, people won't find it. Both positions are correct, which makes the argument mostly unproductive.

The more useful question is not which audience to prioritize. It is where the requirements overlap, where they genuinely conflict, and how to run one content program that serves both without pretending the tension doesn't exist.

The stakes are real. G2's March 2026 survey of 1,076 B2B software buyers found that 51% now start their research with an AI chatbot more often than with Google, up from 29% in G2's 2025 report. Most of them (61%) use AI search alongside Google rather than instead of it. Meanwhile, Ahrefs' 2026 update to its click study found that AI Overviews correlate with a 58% lower click-through rate for the top-ranking page, a modeled estimate based on about 300,000 keywords. The audience has split. Some readers arrive through traditional search. Many now arrive, or never arrive at all, through AI synthesis. A brand that optimizes for only one of these paths is already losing ground on the other.

The good news is that the overlap is larger than most communications leaders assume.

Where Human and Model Requirements Converge

The foundational research on generative engine optimization is a 2024 paper by researchers at Princeton, IIT Delhi, and other institutions, accepted to KDD 2024. They built a benchmark of 10,000 queries, tested nine content optimization methods, and measured which ones improved a source's visibility in AI-generated responses. The top three were citing sources, adding statistics, and including quotations from credible people. Each improved visibility by 30-40% on the paper's position-adjusted word count metric.

Notice what those three techniques have in common. They are also exactly what makes content more credible and persuasive to human readers. A well-sourced argument with specific data and attributed expert opinion is more convincing to a CMO evaluating a vendor than a wall of unsupported claims. The same qualities that make content extractable to a model make it trustworthy to a person.

This convergence runs deeper than formatting.

Specificity rewards both audiences

Generic commentary fails with both readers and models. A human reader skims past "AI is transforming how organizations communicate" because it contains no useful information. A language model deprioritizes it for the same reason: there is nothing specific enough to extract and attribute. A sentence like "AI will restructure demand generation faster than brand strategy because demand generation depends on information routing, while brand strategy depends on human interpretation" gives a reader something to evaluate and a model something to cite.

Authority signals compound across channels

Ahrefs' analysis of 75,000 brands (May 2025) found that branded web mentions had the strongest correlation with brand visibility in Google AI Overviews, at 0.664, well ahead of backlinks at 0.218. These are correlations, not proof of cause, but they point the same way: what independent sources say about a brand matters more than how many sites link to it. Earned media coverage, the traditional PR objective, turns out to be one of the most effective AEO investments available. Third-party validation in credible publications tells both human readers and AI systems that the ideas have cleared an independent editorial standard.

Clarity of structure serves both

AirOps' 2026 State of AI Search report found that pages with sequential headings and a single H1 are cited 2.8 times more often. That same structure makes content easier for human readers to navigate. Descriptive H2 headings that signal what a section actually answers are not a concession to machines. They are good editorial practice that most content skips because writers assume readers will follow them through unbroken prose. Most readers don't. Neither do models.

The practical implication: a significant portion of what makes content effective for human readers also makes it more citable by AI systems. The investment is not duplicated. It is shared.

Where the Requirements Actually Conflict

The overlap is real, but so is the friction. There are places where optimizing for model extraction pulls against what makes content compelling to read.

Front-loading versus narrative build

Human readers are drawn into arguments. A well-constructed essay can withhold its central claim for several paragraphs while building context and tension. That approach works in long-form journalism and executive thought leadership because readers who commit to a piece will follow the reasoning.

AI systems work differently. In Kevin Indig's Growth Memo analysis of 18,012 verified ChatGPT citations, 44.2% came from the first 30% of the content. At least for ChatGPT, models extract answers from the opening of a piece, not from the conclusion of an argument. Content that buries its claim in the seventh paragraph after building to it will be extracted incompletely or not at all.

The conflict is structural. The narrative technique that makes a human reader feel the argument's weight is exactly what causes a model to miss the point.

Resolution: Lead with the claim, then build the argument. A strong declarative sentence at the top of each section satisfies model extraction without eliminating the reasoning that follows. Readers who want the argument can continue. Models get what they need immediately.

Stylistic fluency versus extractable structure

The GEO paper found that improving readability and fluency boosted AI visibility by 15 to 30%. That sounds like alignment. But there is a subtler tension underneath it.

Prose that flows naturally tends to embed information within sentences rather than surfacing it as discrete, attributable facts. A writer might say "the shift toward AI-mediated research has significantly changed how buyers evaluate vendors." A model would prefer: "51% of B2B software buyers now start research with an AI chatbot more often than with Google (G2, 2026)." The second sentence is less elegant. It is also far more likely to be cited.

The instinct to write smooth, connected prose is not wrong. But it needs to be balanced against the discipline of surfacing named statistics, explicit attributions, and discrete claims that a model can lift cleanly.

Depth versus freshness

Thorough, deeply researched content takes time to produce, and keeping it current takes more. AirOps' report found that pages left more than three months without an update are over 3 times more likely to lose citations. The economics of depth and the economics of freshness pull in different directions, and most content programs have not reconciled them.

These conflicts are manageable. But managing them requires acknowledging they exist rather than assuming that good writing is automatically good for models, or that AEO compliance automatically makes content worth reading.

Running One Program That Serves Both

The practical answer is not to run two content tracks. That doubles cost and creates inconsistency. The answer is to build a set of structural habits that satisfy model requirements without degrading the quality of the argument for human readers.

The section architecture that works for both

Each major section should follow a consistent internal structure:

1. Direct answer or claim (first 1-2 sentences). This is what the model extracts. It should be a complete, attributable statement that makes sense in isolation.
2. Evidence (a named statistic, study, or example with explicit attribution). This is what both audiences need to trust the claim.
3. Reasoning (the argument that connects the evidence to the claim). This is what human readers came for.
4. Implication (what the reader should do or think differently as a result). This is what makes the section commercially useful.

Most thought leadership content contains only the third element. It reasons without claiming, evidences without concluding, or concludes without connecting the conclusion to action. Adding the first and fourth elements costs little in word count. The return, for both audiences, is significant.

Treat the introduction as the highest-value real estate

The Growth Memo finding that 44.2% of ChatGPT citations come from the first 30% of content has a clear operational implication: the introduction should carry the highest density of specific, citable information in the entire piece. Not scene-setting. Not throat-clearing about why the topic matters. The most important claim, supported by the strongest data, in the first 150 words.

This also makes the piece more useful to human readers, who make decisions about whether to continue reading within the first few sentences. An introduction that leads with a surprising statistic and a clear argument earns more continued reading than one that builds slowly to its point.

Freshness as editorial discipline, not just technical maintenance

The finding that pages left more than three months without an update are far more likely to lose AI citations is often treated as a technical problem. It is actually an editorial one. Content that is not being updated is usually content that no longer reflects the current state of the argument. The solution is not to change publication dates. It is to maintain a small number of high-quality anchor assets and update them with new evidence, revised claims, and current data on a regular schedule.

This is a different content model from the one most communications teams run. Producing a high volume of short pieces and abandoning them is efficient in the short term and increasingly costly in both SEO and AEO terms. A smaller inventory of deeply maintained assets outperforms a large inventory of stale ones.

The structured data question

Structured data, such as FAQPage and Article schema, tells machines what a page is and what questions it answers. It is not a writing decision. But it is a communications decision. The question of how an organization's content is interpreted by AI systems is not purely technical. It connects directly to brand authority, message control, and competitive positioning. Communications leaders who leave structured data entirely to technical teams are ceding influence over how their content is understood and cited.

As we've written in our field notes on executive thought leadership for AI search, the same qualities that make content useful to people tend to make it easier for machines to interpret. Clarity, specificity, evidence, and attribution are not separate requirements for separate audiences. They are a single standard that the best content has always met.

The Measurement Problem

Most communications teams measure content performance through a lens built for traditional search: page views, organic sessions, keyword rankings, time on page. Those metrics remain useful but increasingly incomplete.

A piece of content that earns a significant share of AI citations in a category may never generate a direct click. The reader gets the answer inside the AI interface. The brand is mentioned. The association is built. None of that shows up in a pageview report.

This creates a measurement gap that most teams have not yet addressed. The metrics that tell you whether content is working for human readers do not tell you whether it is working for model audiences. Tracking AI citation frequency requires different tools and a different reporting cadence.

What to track:

• Citation frequency in AI engines. How often does your brand appear as a cited source when AI systems answer questions in your category? This is the headline AEO metric.
• Share of voice in answers. Of the sources cited by AI engines for your priority queries, what percentage is yours?
• Branded search volume. A downstream signal that shows up in standard analytics. In Ahrefs' 75,000-brand study, branded search volume correlated with brand visibility in AI Overviews at 0.392, so a rise in branded search is worth watching alongside citation data.

The brands that build measurement infrastructure for both audiences now will understand where they stand before their competitors do. That head start compounds, because every month of tracking adds a baseline the late movers will not have.

The Argument Worth Having

The debate between writing for people and writing for models is largely a false choice, but it is not entirely false. There are real tensions, and pretending otherwise leads to content that resolves neither set of requirements well.

The more productive framing is this: what does it mean to produce content that is genuinely useful? Useful to a reader who needs to understand something, make a decision, or evaluate a vendor. Useful to a model that needs a specific, attributable claim it can incorporate into a synthesized answer.

The answer to both questions is the same kind of content: specific rather than generic, evidenced rather than asserted, structured rather than meandering, and maintained rather than abandoned. That is not a new standard. It is the standard that separated good thought leadership from filler before AI systems existed.

What AI systems have done is make the cost of ignoring that standard visible faster. Content that was always vague, unsourced, and structurally incoherent now fails more quickly and more measurably than it did when the only audience was a human reader with limited alternatives.

That is probably a useful development for the quality of communications work overall.

Partner With Us →