Optimize for the question, not the model
Refreshed for current answer engines; the original analysis is preserved.
Google launched its most capable model this month, the benchmarks moved, and for a while there is a new leader. It is exciting, and it is the wrong thing to build a strategy around. Models turn over every few months. The question your buyer asks an AI does not. The whole of a durable GEO strategy sits in that gap.
the short answer
Models churn — a new version, new benchmarks, a new leader every few months — so optimizing for the model of the moment buys work with a short shelf life. What is stable is the question your buyer asks, because it follows human decisions, not model releases. Optimize for that: be the clearest, best-sourced, most verifiable answer to your category’s durable questions, in a form any competent model can read. That work serves the model leading today and the one nobody has announced yet.
key takeaways
- Google launched Gemini this month and the leaderboard moved again. Chasing the model of the moment is a losing strategy, because the model of the moment keeps changing.
- Models churn — new versions, new benchmarks, a new leader every few months. The questions your buyer asks an AI do not churn; they follow human decisions, which are stable.
- So optimize for the question, not the model: be the clearest, best-sourced, most verifiable answer to the durable queries in your category, in a form any competent model can read.
- Deep tuning to one engine’s current behavior is a trap — that behavior is the most temporary thing in the system. The model-agnostic engines of visibility are entity, sources, and corroboration.
- A more powerful model does not change the strategy. It reads good groundwork more faithfully and exposes thin groundwork more plainly. Build for what every model is trying to do.
what moves, what doesn’t
The row on top rearranges itself every few months; the bar underneath is the same question your buyer was asking last year and will ask next year. Tie your work to the bar, not the row, and the churn above stops being your problem.
Why model-chasing feels right and ends wrong
Chasing the model is seductive because it is concrete and current. There is a launch, a number, a leader, a clear thing to react to — and reacting to news feels like strategy. But the very concreteness that makes it feel actionable is what makes it fragile: you are reacting to the single most temporary fact in the field. By the time you have reorganized around the behavior of this quarter’s best model, the field has produced a new best model, and the behavior you tuned for is no longer the behavior that matters. The effort was real; the durability was not. It is the strategic equivalent of building on sand because the sand happened to be where the tide was today.
The deeper problem is opportunity cost. Every hour spent tailoring to a model’s present quirks is an hour not spent on the work that would have helped no matter which model was reading — clarifying your entity, deepening your answers, earning corroboration. Those compound; model-tuning evaporates. A strategist’s job is to notice which efforts accrue and which expire, and to pour the budget into the first. The launch this month is genuinely impressive, and it changes almost nothing about where your effort should go, because the thing it improved — the model — was never the thing you could build a lasting position on.
The argument, in three parts
The model is the fastest-moving thing in the room, the question is the slowest, and the strategy is to build for the second in a form the first can always read. Open each part for where it changes the plan.
01 The model is the fastest-moving thing in the room
Look at the pace. A model launches, tops a benchmark, and is hailed as the new state of the art; a few months later another one passes it, and the cycle repeats, faster each time. This month’s launch — a major new model arriving with leaderboard scores that edge out the previous best — is not an exception, it is the rhythm. For a strategist, the lesson is not which model is ahead today; it is that "ahead today" is a perishable fact. Anything you build that depends on the specific behavior, ranking, or quirk of the current leader inherits that short shelf life. You will have spent real effort optimizing for a configuration of the world that is already scheduled to be replaced. The churn is not a temporary phase before things settle down; rapid turnover is the settled state, and a strategy has to be built to survive it rather than to ride one wave of it. The teams that treat each launch as a moment to react are signing up to react forever, because the launches do not stop; the rhythm only quickens, and a plan that depends on keeping pace with it is a plan that depends on never resting.
02 The question is the slowest-moving thing in the room
Now look at what does not move. A person trying to choose a tool in your category asks, in some form, the same handful of things they asked last year and will ask next year: what is this, which option fits my situation, how do these two compare, can I trust this vendor, what does it cost. Those questions are anchored to human decision-making, not to model architecture, and human decision-making changes on the scale of years and culture, not months and releases. That stability is the most valuable thing a strategist has, because it is the one part of the system worth making a multi-year bet on. If you become the clearest and most credible answer to the durable questions of your category, you are useful to the model that leads today, the one that leads next quarter, and the one nobody has announced yet — because all of them are trying to do the same job, which is to answer those questions well. The model is the variable; the question is the constant, and you build a lasting position on constants, not on variables that were designed from the start to be improved upon and discarded.
03 So build for the question, in a form any model can read
The practical move follows directly. Write down the real questions your buyer asks — in their words, not your keywords — and set out to be the best answer to each: a clear, liftable answer at the top, the depth that makes yours the one worth trusting underneath, and the sourcing and entity clarity that let any engine recognize and rely on you. None of that is tuned to a particular model, which is the point; it works because it reflects what every model is built to do rather than how one of them happens to do it this month. This is the discipline the AC Group has practiced for 27 years under changing names — answer engines, AI search, generative experiences — because the underlying bet never changed: serve the question, stay legible to whatever reads it, and let the models come and go beneath you. We have watched a dozen supposed revolutions in how machines read the web, and the durable response to each was the same: serve the question better than anyone, and stay legible. The technology changed; the discipline did not.
What this looks like in practice
Two teams react to this month’s launch differently. The first treats it as a fire drill: it studies the new model’s habits, rewrites its pages to match what this version seems to favor, and congratulates itself on being early. For a while it may even see a bump. Then the next model arrives with different habits, the rewrites no longer fit, and the team is back at the start, having earned a result with the lifespan of a season. It mistook motion for progress, and the motion was tied to the one thing guaranteed to move on.
The second team barely changes course. It reads the launch as confirmation that the engines are getting better at the job they were always doing — answering questions from sources — and it keeps doing the work that serves that job: making sure its entity is unmistakable, its answers to the category’s real questions are the clearest and best-sourced available, and its story is corroborated where credible sources can be found. When the new model reads the web, it finds that team well-prepared, not because they guessed its preferences but because they built for what it, and every model, is fundamentally trying to do. A year and three model launches later, the second team is compounding and the first is still reacting, having paid three times for results that each lasted a season while the second paid once for a position that kept appreciating. That gap is the whole case for optimizing the question.
What to do with this
Make a list — the real one. Write the questions a prospect actually asks when they are trying to solve the problem you solve, in the language they would use, not the keyword fragments a tool suggests. What is this category and do I need it; which option fits a situation like mine; how do the leading choices compare; can I trust this vendor; what will it cost and what do I get. That list is your strategy document, because it names the durable surface you are competing on, and it will look almost the same a year from now when three new models have come and gone.
Then do the model-agnostic work against that list, and resist the urge to do anything model-specific beyond basic hygiene. Be the best answer to each question: clear at the top so an engine can lift it, deep underneath so it is the one worth trusting, sourced and entity-clear so any model can recognize and cite you. Keep it current, because a stale answer to a stable question is its own kind of failure. And measure yourself on whether you are becoming the answer to those questions across engines, not on whether you cracked the behavior of the model that happens to lead this month. The leader will change. The question will not. Build for the one that stays, and let the launches — this one and the dozen after it — wash past a position they cannot move, the way they have for the ' + years + ' years we have done this under one banner or another.
A test for any GEO tactic
There is a simple question that sorts durable work from disposable work, and it is worth applying to every tactic before you fund it: would this still help if the leading model were replaced tomorrow? If the answer is yes — clearer pages, a cleaner entity, better sources, a more complete answer to a real question — the work is an asset, and it compounds. If the answer is no, because it depends on a behavior, a phrasing trick, or a formatting quirk that the current model happens to reward, then you are not building an asset; you are renting a result from a landlord who changes the locks every quarter. The test costs nothing to run and saves a great deal, because most of what gets sold as an AI-visibility tactic fails it quietly. The vendors selling the failing tactics are not always cynical; they are often just optimizing for what looks impressive in a demo this quarter, which is exactly the thing the test is designed to see through.
Run the test honestly and the roadmap almost writes itself. The things that pass — entity clarity, current and well-sourced answers, corroboration, plain structure a machine can parse — are exactly the things that are tedious to do and easy to defer, which is why so many teams skip them for the quick win that fails the test. The discipline is not cleverness; it is refusing the tactic that feels current in favor of the one that stays true. A strategist earns their keep less by spotting the next trend than by declining to chase it when it will not survive contact with the next model.
Why this is good news, especially for smaller players
If keeping up meant tracking every model and re-tuning for each release, the advantage would go to whoever has the most people to throw at the churn — the largest teams, the biggest budgets. Optimizing for the question inverts that. The question is stable, so the work against it is done once and maintained, not redone with every launch, and a focused team that has genuinely become the best answer to its category’s real questions holds that position across models without an arms race it cannot win. You do not need to out-spend the field on chasing models; you need to out-think it on serving the question, which is a contest of clarity and care rather than of headcount.
That is why the model-agnostic path is not just more durable but more democratic. It rewards the team that understands its buyer well enough to name their real questions and answer them better than anyone, regardless of size. The launch this month, and every launch after it, raises the ceiling on what the engines can do with good groundwork — which quietly favors whoever did the groundwork. The smaller player who built for the question gets the same upgrade as everyone else, for free, every time a better model ships. That is the opposite of an arms race; it is a tide that lifts the prepared. And preparation, unlike a model advantage, does not expire when someone ships something bigger; it is the rare kind of work whose value goes up, not down, as the technology around it improves.
Question, not model: quick answers
What does it mean to optimize for the question instead of the model?
It means building your visibility around the questions your buyers actually ask an AI, rather than around the quirks of whichever model is winning benchmarks this quarter. Models change constantly — a new version, a new context window, a new leader on the leaderboard every few months — and any work tuned to one model’s particular behavior ages out when the next one ships. The questions, by contrast, are stable: people keep asking what a category does, which tool fits their case, how two options compare, and whether a vendor can be trusted. Optimizing for the question means being the clearest, best-sourced, most verifiable answer to those durable queries, in a form any competent model can read. You are building for the part of the system that does not churn.
Should I tune my content for a specific AI engine?
Be careful with it. Light, principled adjustments are fine — making sure your pages are crawlable, clearly structured, and accurate helps every engine and hurts none. But deep tuning to one engine’s present behavior is a trap, because that behavior is the most temporary thing in the system. The model you optimized for gets replaced, retrained, or rebalanced, and the trick that worked stops working or quietly backfires. The durable engines of visibility are not engine-specific: a clear entity, current and well-sourced answers to your category’s real questions, and corroboration across credible third-party sources. Those help every model because they reflect what every model is trying to do — find a trustworthy answer — rather than the accident of how one of them does it this month.
How do I know which questions to optimize for?
Start from your buyer, not your keyword list. The questions that matter are the ones a real prospect would type or speak when they are trying to solve the problem you solve: what is this category, which option fits a situation like mine, how does A compare to B, is this vendor credible, what does it cost. These are stable because they follow human decision-making, which does not change at the pace of model releases. List them in plain language, the way a person would actually ask, and then be the best answer to each — clear at the top, deep underneath, and sourced well enough that an engine can trust and cite you. The keyword era trained people to think in fragments; the question era rewards thinking in whole, answerable questions a person would recognize as their own.
Does a more powerful model change my strategy?
Less than the headlines suggest. A more powerful model reasons better over what it can find, but it is still trying to answer a question from sources, and it still rewards being the clear, credible, retrievable one. A stronger model does not invent good information about you that was never there; it just handles whatever exists more competently. So the arrival of a more capable model is not a reason to change your strategy — it is a reason to make sure the durable work is done, because a better model will read good groundwork more faithfully and expose thin groundwork more plainly. The strategy is stable precisely because it targets what every model, weak or strong, is built to do.
A note on sources and timing
This is written in December 2023, the week Google announced Gemini, its new multimodal model, and began using it to power Bard — the launch that prompted these notes and a fair example of the leaderboard churn the piece is about. We have used it as an illustration of the pace at which models turn over, not as a verdict on which is best; that verdict, by the argument here, is the perishable part. We have not described capabilities, renames, or products that did not exist as of this writing. The durable claim does not lean on any one launch: models change far faster than the questions buyers ask, so a visibility strategy should target the question and stay readable to whatever model answers it. That is the work the AC Group has done for 27 years, through every banner the field has worn — and the launch this month, impressive as it is, changes none of it.