Same Prompt, Different Answer: What AI Variance Means for a Brand
Run the same prompt twice. “Write a three-sentence product description for Northwind, a scheduling tool for field service teams, in a confident, plain voice.” The first run leads with the outcome. The second leads with the audience. One says “field teams,” the other “crews.” One mentions the mobile app; the other does not, because the prompt did not, and the model decided differently about what mattered.
Neither is wrong. That is what makes this hard. Marketers come to these tools with the mental model of software: same input, same output, and if the output is wrong, fix the input. Language models do not work that way, and the teams that get consistent results are not the ones with the best prompts. They are the ones who stopped expecting the phrasing to be consistent and made everything else structured instead.
This post explains where the variance comes from, why turning the temperature down does not remove it, what it does to a brand at fifty pieces a month, a test to run before trusting any prompt, and the rule that fixes it.
Where the variance comes from
A language model produces text by predicting a probability distribution over the next word and then choosing one. The choosing step is sampling: the most likely word is usually picked, but not always, and the setting that controls how often the runner-up wins is called temperature. Higher temperature, more surprise; lower, less. This is the part most explanations stop at, and it leads to the wrong conclusion, which is that temperature zero means determinism.
It does not, for two reasons the model labs themselves have documented.
Close calls flip on tiny numerical differences. In a large deployment, your request is processed in a batch with other people’s requests, and the size and composition of that batch changes the floating-point arithmetic in ways that are individually negligible and occasionally decisive. When two candidate words are nearly tied, the tie can break differently on Tuesday than it did on Monday. Engineers at one model lab published a detailed account in 2025 of exactly this batch-invariance problem and how much work it takes to eliminate. Consumer products have not done that work, because for a chat assistant it does not matter.
Routing varies. Many current models are mixtures of experts: the request is routed to different internal sub-networks depending on the input and, again, on batching. Different route, slightly different output.
The empirical result matches. A study of code generation found that the same prompt, at temperature zero, produced dozens of distinct outputs across a thousand runs. And most marketers never touch temperature anyway; the consumer chat products do not expose it. Every draft you get is one draw from a range. There is no setting that makes it the answer.
What variance does to a brand
For a one-off, variance is harmless or even useful; you regenerate until you like one. A brand is not a one-off. A brand is the promise that the next thing the company says will feel like the last thing, and a marketing team at a $5M to $100M company ships fifty to a few hundred pieces a month. If each is a draw from a distribution, here is what wanders with the phrasing.
- Facts. One email says the trial is fourteen days, the next says two weeks, the third says “a free trial.” Pricing is “from $249” and then “starting at around $250.” The model was never told these were fixed, so it treated them like wording.
- Voice. The rules held in the first three posts and drifted in the fourth, which used an exclamation mark and “we’re thrilled.” The brand voice drift post covers how this happens inside a single long thread; across fifty separate runs it happens by default.
- Claims. A result that was hedged on Monday is stated flatly on Thursday. The substantiation standard did not change; the draw did.
- The offer. The call to action is “book a demo,” then “start your trial,” then “talk to us,” in three pieces of the same campaign, because the prompt said “include a CTA” and the model chose.
- Visuals. The same thing happens in image models, and the campaign consistency post handles it: every generation is a fresh draw, and chaining from outputs compounds the drift.
The audience does not see variance. It sees a company that is not sure what it is.
The five-run test. Before any prompt goes into a repeatable workflow (the weekly newsletter, the nurture emails, the social captions), run it five times in fresh sessions and put the outputs side by side. If the facts, the voice, the claims or the offer differ, the prompt is not a process; it is a lottery. Fix the inputs before shipping the workflow. The same test exposes sycophantic reviewers: five verdicts on one draft that disagree are a mood, not a review.
The rule: pin the invariants, free the phrasing
The mistake is trying to make the model deterministic. It cannot be, and the effort produces enormous prompts that still vary. The fix accepts variance where it is harmless and removes it where it is not.
Decide what may vary and what may not. Wording, sentence order, the specific example, the transition: those may vary, and that variation is what makes the fiftieth email not read like the first. Facts, voice rules, visual system, claims, the offer and the call to action may not vary. Those are the invariants.
Pin the invariants as structured inputs, not prompt words. This is the whole method. “Our trial is 14 days” in a prompt is a suggestion the model will paraphrase. A facts object attached to every run, from which the model must take the trial length, is a constraint. The difference is between describing your brand to the model and handing it your brand.
| Invariant | As a prompt (weak) | As a structured input (pinned) |
|---|---|---|
| Facts | ”We offer a 14-day trial starting at $249” | A maintained facts file attached to every run; the review flags any fact not in it |
| Voice | ”Confident, plain, no exclamation marks” | Voice rules as a checklist the draft is scored against, with examples of pass and fail |
| Visual system | ”Use our brand colors” | Hex values, fonts and the logo file applied after generation as assets |
| Claims | ”Don’t overpromise” | An approved-claims list; anything outside it is flagged before a person reads |
| Offer and CTA | ”Include a call to action” | The campaign’s one CTA, as a field the draft must use verbatim |
Once the invariants are pinned, the phrasing can vary freely and the brand does not move, because nothing the audience uses to recognize you was left to the draw. The brand voice document is the human-readable version of the voice row; the point here is that it has to become an input the system enforces, not a PDF the model is asked to remember.
Put a review pass after the run that checks the invariants held. Even pinned, a model will occasionally paraphrase a fact or slip the voice. The review is not “is this good”; it is “did the invariants hold,” item by item, with the line quoted. That is a check a machine can run before a person reads, and it is the step that turns a model that varies into a process that does not.
What this looks like built
Marqeable is designed around pinned invariants. Brand voice, brand look and business facts live with the account and are attached to every draft and every image as structured context rather than prompt text; the campaign’s offer and call to action are fields the pieces inherit; the review pass checks each draft against the voice rules, the facts and the approved claims and flags the line that drifted; and a person approves before anything ships. The model’s phrasing varies every run, which is what you want. The trial is fourteen days every time.
Frequently asked questions
Does ChatGPT give the same answer every time?
No. Models sample the next word from a distribution, so runs diverge, and batching and routing in large deployments add variation even at the lowest settings. Every output is one draw from a range.
Does temperature zero make it deterministic?
It reduces variation without eliminating it; studies found dozens of distinct outputs across a thousand runs at temperature zero, and model labs have documented the batch effects behind that. Consumer chat products do not expose temperature anyway.
Why does this matter for a brand?
Because a brand is the promise of consistency, and at fifty pieces a month the facts, voice, claims and offer wander with the phrasing unless something pins them.
How do you get consistent output from a non-deterministic model?
Pin the invariants (facts, voice, visuals, claims, offer) as structured inputs attached to every run, let only the wording vary, run the five-run test before trusting a prompt, and check the invariants in a review pass after.
The bottom line
A model is not software. Same prompt, different answer is not a bug to be prompted away or a setting to be turned down; it is what generation is. A brand cannot survive fifty draws a month unless the things that make it recognizable are taken out of the draw. Decide what may vary, pin what may not as inputs the system enforces, test every repeatable prompt five times, and review for whether the invariants held rather than whether the draft felt good. Then let the phrasing vary. That part was never the brand.
Marqeable runs your campaigns, answers every visitor, text, and email in seconds, and turns them into booked jobs and meetings - even at 9pm on a Saturday. We’re in private beta with a small early cohort. Get early access
