How to test AI in production: criteria and metrics

You run marketing, SEO or a B2B editorial line, and an AI already writes part of your pages: articles, product sheets, FAQs, summaries. Here, testing an AI in production means checking the one that writes for you, before its outputs reach Google and generative engines.

This page helps you decide what to test, against which criteria, who signs off and what blocks publication. Checking texts that have already been produced belongs to AI content detection, covered separately.

What you test when an AI writes for you

Testing a content AI means verifying that it produces useful, stable, compliant outputs in conditions close to production. You do not test a single answer: you test behaviour in context, with a brief, instructions, allowed sources and a usage charter. The system under test includes the model, the prompts, the templates and the data supplied.

A valid B2B test stays traceable (who approved what), reproducible (same inputs, same outputs or explainable differences) and auditable (logs and versions).

Define the scope before the first run

Formalise four elements before comparing any versions:

• Use cases: generating articles, pages, FAQs, summaries or briefs.
• Constraints: tone, length, structure, mandatory fields, forbidden outputs, required level of evidence.
• Allowed data: internal sources, validated documents, reference pages the system can consult.
• Stop criteria: recurring invented claims, non-compliance with the charter, unstable structure.

Without this frame, your teams keep "impressive" outputs that cannot be used in production.

Build the test set: real briefs and edge cases

A strong test set mixes real scenarios (80%) with edge cases (20%). Real scenarios reuse your briefs, your search intents, your customers' questions and your brand constraints. Edge cases put the system under stress.

Three families of edge cases reveal the weaknesses of a content AI early:

• Ambiguous brief: the intent is unclear, and the AI should ask for clarification rather than choose on its own.
• Conflicting instructions: length versus completeness, tone versus format, and the AI should make a readable trade-off.
• Missing context: a data point or source is not supplied, and the AI should flag the gap instead of inventing.

For GEO, add conversational requests: short questions, multi-question prompts, comparisons, summaries, definitions backed by evidence. Attach a verifiable expectation to each case: a structure, a definition, a cited source or an admission of uncertainty.

Avoid measurement bias

Three mistakes skew a test before you even read the results:

• Data leakage: test on examples absent from internal briefs and from the examples given to the model.
• Unstable prompts: lock the prompt and temperature, then change only one variable at a time.
• Sample too easy: avoid similar examples, narrow vocabulary and identical structures.

A sign of maturity: your tests surface failures early, not just successes late.

The criteria

Criteria turn your definition of quality into controllable checkpoints. For an AI that writes, they cover the accuracy of facts and figures, sources, compliance with the brief and the charter, readability and quotability. Robustness comes on top: a criterion met once but not under a rephrased instruction is worthless in production.

The following table pairs each criterion with a test and a qualitative acceptance threshold, to be set according to what is at stake on the page.

Quality is therefore not a single score: it depends on the page type (article, category, product) and the search intent it serves.

Accuracy of facts and figures

Accuracy means no unsubstantiated claims and factual consistency from one end of the text to the other. For figures, the test is binary: every value points to a supplied source, with its year. An AI that rounds a value or attributes a figure to the wrong source fails, even if the text reads well.

Compliance with the brief and the charter

The check covers elements that can be verified line by line, which your internal scoring rules can partly automate:

• Expected structure: headings, sections and order set by the template.
• Mandatory elements: definition, example or statement required by the brief.
• Tone and register: in line with the brand charter, without unnecessary jargon.
• Forbidden items: no term, topic or wording excluded by the charter.
• Intent: the output answers the reader's question, not a neighbouring one.

These checks spot deviations; human review decides on the pages that matter.

Robustness: the same quality when the instruction varies

Test robustness by varying the instruction order, synonyms, an extra constraint or partial context. Measure whether outputs stay compliant and whether the AI asks for clarification rather than inventing. For GEO and SEO, robustness also shows in structural stability: does the AI keep coherent headings and stable definitions when the phrasing changes?

Readability and engine-oriented quality

Four elements make a passage usable by an engine as well as by a reader:

• Entities: business terms, products and concepts correctly defined.
• Evidence: data, limits and conditions, with no unsupported statements.
• Definitions: clear opening sentences that are useful for excerpts.
• Freshness: dated updates when they change the answer.

Test the model's ability to produce short, self-contained passages that combine definition, context and nuance: this is typically what engines reuse. The criteria for judging a text once it has been produced are covered in our article on the quality of AI-generated text.

Quotability tests: ask, cite, state uncertainty

Create scenarios where the answer calls for a source: a statistic, a rule, a sensitive recommendation. The test checks that the AI adopts one of the three acceptable behaviours:

• Ask for clarification when the instruction does not allow a correct answer.
• Cite a supplied source without fabricating another one.
• State uncertainty instead of filling the gap with a claim.

Do not force citations everywhere: in SEO, unnecessary outbound links hurt the experience. In GEO, the goal is verifiability on critical points, not a systematic bibliography.

GEO prompt scenarios

Replay requests close to those generative engines receive, and read what the AI does with them:

• "Explain X in 5 points, with a definition and 2 limitations."
• "Compare X and Y using 3 criteria, then recommend based on context."
• "Provide an operational checklist and the common mistakes."
• "Answer like a B2B expert, without unnecessary jargon."

Your AI passes if the answer stays precise and structured, and does not flatten nuance. That is often what separates a merely plausible response from a genuinely usable one.

Reproducibility and non-regression

A repeatable protocol lets you compare versions of prompts, models or templates without endless debate. You are looking for an operational truth: in our use cases, does this version produce better content, with less risk? Keep one principle: if you cannot replay a test, you cannot trust its result.

The non-regression core

Build a stable core of invariant tests, replayed at every change of model, prompt or template. Then add tests specific to a given change: a new language, a new SEO format. Always include tests for definition, evidence, structure and the ability to stay factual: that is where regressions hide when you push for creativity.

Each type of change calls for a specific test before going into production:

After a fix, replay the exact same tests: without a non-regression loop, the corrected defect often returns in the next version.

Version, trace, replay

Four elements make a test replayable and auditable:

• Versions: identifiers for the prompt, model, template, data and rules.
• Logs: inputs, outputs, errors and context metadata.
• Test set: real scenarios and edge cases, with documented expectations.
• Reports: deltas by criterion, page type, language and version.

For each run, keep the results by criterion and the decision taken, approved or rejected, with its rationale. This traceability lets you explain a regression, prove an improvement and secure a rollout across several sites or markets.

Automate the repetition, not the decision

Automation exists to industrialise repetition, not to replace judgement. A useful testing pipeline rests on four settings:

• Scheduled runs: a regular pass on the core, weekly for example.
• Triggered runs: a full pass at every change of model, prompt or template.
• Alert thresholds: human review as soon as a score drops or errors increase.
• Decision-oriented reports: high-risk cases first, to fix, roll back to the previous version or approve.

Human validation and acceptance thresholds

Automated scores flag deviations; human review decides. The AI tester, or the lead who plays that role, turns editorial objectives into testable criteria, organises cross-validation between business, data and SEO, then documents decisions.

Each role carries part of the decision, at a specific point in the cycle:

The scoring rubric and dual review

Create a 5- or 7-point rubric per criterion: accuracy, coverage, clarity, compliance with the charter, sources. Run dual reviews on a subset of the test set, then measure agreement between reviewers. Frequent disagreement signals a rubric that is too vague: sharpen the criteria before judging the model.

Which sample to review

Exhaustive review is expensive and does not hold at scale. Instead, place checkpoints where risk concentrates:

• Before publishing: a sample of high-traffic or high-stakes pages.
• After publishing: pages showing abnormal signals, such as a falling CTR or a high exit rate.
• On alert: any significant non-regression deviation triggers a read-through.
• On a major change: a new model or a new prompt goes through a full review of the core.

What blocks publication

A good average score does not protect you from a critical case. An output does not go live, whatever its overall score, in the following cases:

• a claim or figure without a verifiable source;
• a charter prohibition crossed;
• a passage that loses its meaning when read on its own;
• an unexplained regression on the core.

The acceptance threshold is set by what is at stake on the page, not by a score: a high-traffic or high-stakes page goes through human review, even when every indicator looks good.

Measure the editorial effect

Compare by content batch and by page type, before and after a model or prompt change: impressions, clicks and positions in Search Console, engagement and conversions in Analytics. The post-publication measurement method is detailed in our article on how to measure the reliability of an AI detector.

Testing AI and detecting AI: two different moves

Testing an AI validates the system before release, whereas detection examines an already-produced text to spot signals of generation. The two complement each other without overlapping: over-investing in downstream checks means treating effects rather than upstream causes. The methods are explained in our article on how to detect text generated by artificial intelligence, and the protocol applied to a text in our guide to check whether a text was generated by AI.

Depending on your situation

Depending on the stage of your content chain, these articles take over:

• You are building the test set from your queries: start from search intent to choose the real scenarios.
• You write the briefs the AI receives: rely on our SEO copywriting method to set angle, structure and mandatory elements.
• You set the charter the AI must follow: tie it to your editorial strategy so that tests measure what matters.
• You want generative engines to reuse your content: our guide to generative engine optimization details what makes a page quotable.

If your challenge is to test an AI that already writes with your charter and your sources, Incremys offers a personalised AI that frames generation upstream, with traced versions and approvals.

AI Testing FAQ

What is AI testing?

AI testing covers the human and automated checks that demonstrate an AI produces useful, stable outputs that comply with its intended use. For an AI that writes your content, it covers accuracy, sources, compliance with the brief and the charter, readability, quotability and stability from one version to the next.

How do you test an AI system?

You test an AI by first defining the scope, then running it against a test set made of real briefs and edge cases. Each output is scored on explicit criteria, with dual review on a sample. Versions are compared on the same set, and every decision is documented before moving into production.

What are the key criteria for evaluating an AI system?

The evaluation criteria for a content AI are the accuracy of facts and figures, sources, compliance with the brief and the charter, readability and quotability. Robustness completes them: a criterion must hold when the instruction is rephrased. Each criterion gets a test and an acceptance threshold set according to what is at stake on the page.

How do you measure an AI system's performance?

A content AI's performance is measured first on output quality, with a human scoring rubric and automated checks replayed at every version. You then connect these results to the real editorial effect, comparing content batches before and after a change in Search Console and Analytics.

What is the difference between testing an AI and detecting AI-generated content?

Testing an AI validates the system's behaviour before release, whereas detection analyses an already-produced text to spot signals of generation. Testing acts on causes, upstream; detection serves as a downstream check. The two complement each other, but detection does not replace testing.

How do you design a test set that reflects your SEO intents and GEO use cases?

A useful test set turns your SEO intents, whether informational, comparative or transactional, into realistic briefs. You then add conversational requests for GEO (summaries, comparisons, checklists) and edge cases. Each test gets a verifiable expectation: structure, definition, evidence or source.

Which ongoing metrics help you avoid regressions after a model update?

After a model update, continuously track the quality rubric scores, the factual error rate and output variability on the non-regression core, with alerts on deviations. For SEO, monitor impressions, clicks, CTR and positions; for business, engagement and conversions.

When should you require human validation even if performance scores look good?

Human validation is required on high-traffic or high-stakes pages, at every major change of model or prompt, and as soon as a warning signal appears: a non-regression deviation, an increase in errors, a drop in Search Console. A good average score does not protect you from a critical case.