Build an AI visibility baseline by freezing Cycle 0 before changing content, technical evidence, digital PR, comparison positioning, or source outreach. Define the category, buyer context, prompt groups, competitors, platforms, modes, market, language, labels, and denominators first. Then capture repeated prompt-platform runs, measure the starting noise, and archive the evidence against which later cycles will be compared.
The baseline is neither a single visibility score nor the best-looking answer from a manual check. It is a controlled before-state. Its job is to show what ChatGPT, Gemini, Perplexity, Google AI Overviews, Google AI Mode, or another relevant surface returned under declared conditions—and whether that evidence is stable enough to justify optimization.
The Short Answer: Freeze Cycle 0 Before You Optimize
Cycle 0 is the first locked measurement cycle. It should preserve both the observed visibility and the amount of answer volatility around it. During that cycle, do not revise prompts, replace competitors, switch answer modes, change labels, or start the work whose effect you intend to measure.
A useful baseline produces four outputs:
| Baseline output | What it must contain | What it decides |
|---|---|---|
| Measurement contract | Scope, prompt versions, platforms, modes, market, language, competitors, capture window, labels, and denominators | Whether later data will be comparable |
| Evidence archive | Raw answers or sufficient excerpts, visible sources, answer format, capture time, and classification notes | Whether another reviewer can audit each label |
| Starting-noise profile | Repeated-run distributions for important prompt-platform conditions | Whether the result is stable, mixed, or too volatile to call |
| Action gate | Criteria for choosing whether to optimize, monitor, rerun, inspect sources, audit accuracy, or rebuild the setup | Whether a finding is ready for work |
The freeze applies to the evidence environment, not only the dashboard. Rewriting a category page, publishing a comparison, correcting product facts, starting outreach, or changing how the brand is described can all break the before-state if they happen before Cycle 0 is complete.
Decision rule: if another reviewer cannot reproduce the scope and understand why each row received its labels, optimization is premature.
Define the Baseline Scope and the Decision It Must Support
Start with the decision, not the metric. “Measure AI visibility” is too broad. A usable baseline question is: “Across the declared buyer-intent prompts and answer surfaces, is the brand present, recommended, accurately described, and supported by visible sources relative to the declared competitors?”
That question tells the team what belongs in the baseline and what does not.
| Scope field | What to freeze | Risk if it remains vague |
|---|---|---|
| Business decision | The action the baseline may trigger | Data accumulates without changing a decision |
| Category | The product category or use case in which the brand should reasonably compete | Absence is treated as failure in irrelevant topics |
| Audience | The buyer, evaluator, operator, or researcher represented by the prompts | Generic prompts return a mixed consideration set |
| Market and language | Country, region, language, and local buying context where relevant | Local sources and competitors are blended |
| Brand identity | Brand name, product names, and approved aliases used for entity matching | Mentions are missed or unrelated terms are counted |
| Capture window | The start, end, and collection pattern for Cycle 0 | One busy day is compared with a longer later period |
| Platform condition | Answer engine, product surface, mode, session state, and source visibility | Unlike answer environments are averaged together |
| Classification rules | Definitions for presence, recommendation, citation, position, accuracy, and sentiment | Reviewer judgment changes become apparent movement |
| Denominator | Runs, prompts, eligible recommendation answers, ordered answers, or source-visible answers | Percentages look precise but cannot be interpreted |
The category boundary needs its own check before prompts enter the panel.
| Scope class | Meaning | Baseline treatment |
|---|---|---|
| Core | The brand genuinely serves the buyer need and should be considered | Include in the primary baseline |
| Adjacent | The brand may be relevant, but absence is not automatically a failure | Track separately as exploratory or secondary evidence |
| Out of scope | The product does not solve the tested need or serve the stated market | Exclude from visibility KPIs |
Use this sequence to lock the scope:
- Write the business decision the baseline must support.
- Define the core category, buyer, market, and language.
- Mark candidate prompts as core, adjacent, or out of scope.
- Select only answer surfaces that the audience plausibly uses and the team can capture consistently.
- Write the labels and denominators before collecting answers.
- Set the Cycle 0 observation window and prohibit measurement changes until it closes.
Practical check: if stakeholders disagree about the category, market, or intended buyer, run an exploratory scan instead of publishing a formal baseline percentage.
Build Prompt Groups That Represent Buyer Intent
A baseline should cover decisions across a topic, not reward the brand for a few favorable questions. There is no universal prompt quota. If the panel still needs design work, first build a stable prompt set for AI rank tracking. The right panel is large enough to represent the important buyer intents and small enough that every answer can be reviewed, classified, and repeated under stable conditions.
Separate the panel into prompt groups because each group tests a different type of visibility.
| Prompt group | What it tests | Decision it supports |
|---|---|---|
| Category discovery | Whether the brand appears before a buyer names any vendor | Is the brand discoverable in the category? |
| Problem-aware | Whether the answer connects a buyer problem with the category and relevant options | Does the brand enter consideration from the problem stage? |
| Alternatives | Whether the brand appears when a buyer seeks a substitute for another solution | Is the brand recognized as a credible alternative? |
| Comparison | How named options are differentiated on a specific use case or constraint | Is the comparison accurate and competitive? |
| Recommendation | Whether the brand is selected, shortlisted, caveated, or rejected for a buyer scenario | Does visibility become consideration? |
| Branded validation | Whether the answer understands the named brand, product, capabilities, and limitations | Is the representation accurate and current? |
| Source-sensitive | Which visible pages, domains, or source types support the answer | Which evidence layer needs inspection? |
Do not blend branded validation with discovery. When a prompt supplies the brand name, a mention mainly shows recognition after the entity has been provided. That can expose inaccurate positioning or outdated facts, but it does not prove that an unprompted buyer would discover the brand.
Before a prompt enters Cycle 0, require four conditions:
- A real buyer, evaluator, marketer, or operator could plausibly ask it.
- The tracked brand is genuinely in scope for the need expressed.
- The prompt belongs to one primary group and has a fixed version.
- Its result can lead to monitor, rerun, inspect sources, audit accuracy, or optimize.
Prompt variants should not be used to inflate sample size. If two questions express the same intent with minor wording changes, either choose one stable version or treat the variants as a declared sensitivity test. Silent rewrites turn prompt drift into apparent visibility movement.
Red flag: a panel dominated by questions that name the brand may produce strong recognition numbers while hiding weak category discovery.
Lock Competitors and the Platform Matrix
Competitive measurement is valid only when the comparison set is declared before collection. Start with competitors that a buyer could realistically consider in the same category and context, and keep unexpected brands in a separate observed-competitor field. If the set itself is disputed, use the stricter process for choosing competitors for AI brand tracking before locking Cycle 0.
| Competitor class | How to use it | Main risk |
|---|---|---|
| Direct competitors | Include in the fixed comparison set when they solve the same core problem for a similar buyer | Excluding a material rival makes share comparisons incomplete |
| Category leaders | Include when they define the shortlist even if their scope is broader | They can dominate broad prompts that need tighter intent |
| Adjacent alternatives | Track separately when buyers may use them for part of the same job | Mixing them with direct rivals can create false losses |
| Observed competitors | Record brands that appear unexpectedly during collection | Adding them to the denominator mid-cycle invalidates the original comparison |
An observed competitor can become a declared competitor in the next version of the baseline. It should not be inserted retroactively because its appearance was surprising. If the competitor set changes, label the change and start a new comparison line.
The same discipline applies when tracking brand visibility across AI engines. ChatGPT, Gemini, Perplexity, Google AI Overviews, and Google AI Mode are separate answer surfaces, not interchangeable repetitions of one engine. Their modes, visible sources, localization, session context, and answer formats can differ.
| Surface | Row condition to record | What not to blend silently |
|---|---|---|
| ChatGPT | Search-enabled or model-only mode, clean or continuing session, locale, and visible-source status | Search answers with model-only answers |
| Gemini | Gemini app mode, session condition, locale, and visible links | Gemini app with Google search surfaces |
| Perplexity | Search mode, source visibility, locale, and answer format | Citation-rich answers with no-source observations from other tools |
| Google AI Overviews | Query, market, language, device context where relevant, visible sources, and answer presence | AI Overviews with Google AI Mode |
| Google AI Mode | Mode, market, language, follow-up context, and visible sources | First-turn checks with contextual follow-ups |
These are candidate surfaces, not a compulsory list. A focused baseline across relevant, repeatable surfaces is more defensible than broad coverage that cannot be repeated or acted on.
The basic unit should be explicit:
One exact prompt × one platform × one mode × one market-language condition × one run.
Decision rule: select platforms by audience relevance and repeatability. Segment them before calculating any cross-platform summary.
Capture One Auditable Row per Prompt-Platform Run
Do not begin with a dashboard total. Preserve the evidence row first, then calculate summaries from consistent rows. Screenshots can support review, but they do not replace structured fields.
| Baseline field | What to record |
|---|---|
| Prompt ID and version | Stable identifiers for the exact wording |
| Prompt group and scope | Buyer-intent group plus core, adjacent, or out-of-scope status |
| Platform and mode | Exact answer surface and declared operating condition |
| Market and language | Localization context or a clear not-applicable label |
| Capture time | Date and time of the run within the Cycle 0 window |
| Answer format | Ordered list, unordered list, table, paragraph, source panel, hybrid, or no relevant answer |
| Raw evidence | Full answer where permitted, or a sufficient excerpt for review |
| Visible sources | URLs or domains shown by the surface, plus source type when useful |
| Competitors | Declared competitors and observed competitors in separate fields |
| Classification labels | Presence, recommendation, prominence, citation, sentiment, accuracy, and confidence |
| Denominator eligibility | Whether the row belongs in mention, recommendation, position, or citation calculations |
| Action note | Monitor, rerun, inspect sources, audit accuracy, optimize, or exclude |
Keep the component signals separate:
| Signal | Valid label examples | What it must not imply |
|---|---|---|
| Presence | Absent, named, shortlisted, or unclear | A mention is not automatically a recommendation |
| Recommendation | Selected, favored, neutral, caveated, dismissed, or not applicable | A neutral list entry is not a win |
| Position or prominence | Numeric position in an ordered list, table placement, supporting text, or no valid rank | Paragraph order is not always a ranking |
| Citation | Own domain, third party, competitor, no visible source, or not applicable | A visible citation does not reveal the full hidden source path |
| Accuracy | Accurate, incomplete, outdated, misleading, unsupported, or unclear | Positive tone does not prove factual accuracy |
| Sentiment | Favorable, neutral, caveated, or negative | Tone should not replace a factual review |
Every rate needs the correct denominator. Mention coverage can use all valid in-scope prompt-platform runs. Recommendation rate should use prompts where a recommendation is possible. Position should use only answers with a genuinely ordered format. Citation rate should use source-visible answers, not a mixture that includes modes where sources were unavailable.
For example, “40% visibility” is not interpretable until the report states whether it means two of five runs, forty of one hundred prompts, source-visible answers only, or a weighted composite. Prefer x-of-n reporting at the row and segment level before creating a summary score.
Decision rule: if the metric cannot point back to exact prompts, platform conditions, evidence, labels, and its denominator, it is not ready to drive optimization.
Measure Starting Noise Before Setting the Baseline
AI answers are probabilistic, so Cycle 0 needs a noise read as well as a point value. Repeated runs do not guarantee truth. They show whether the tested condition produces a recognizable pattern or rotates too much for the intended decision. The detailed decision-risk method explains how many AI tracking runs create a clear signal without treating any count as a guarantee.
Use the run count as a risk ladder:
| Run count | What it can support | What it cannot support |
|---|---|---|
| 1 run | A snapshot, evidence example, accuracy issue, or reason to investigate | A trend, stable rank, or confident visibility rate |
| 3 runs | A quick volatility check for a lower-risk prompt | Strong confidence or a major optimization decision |
| 5 runs | A cautious operational read for an important prompt | Proof of causation or a guaranteed future outcome |
| 10 or more runs | A stronger distribution for volatile, high-value, or high-stakes conditions | A repair for a biased prompt panel or mixed platform setup |
This ladder is decision guidance, not a universal statistical guarantee. Some prompts stabilize quickly. Others continue to rotate brands, sources, and answer formats. The report should preserve that distribution rather than hide it inside one average.
An illustrative row summary could say:
- The brand appeared in 4 of 5 runs.
- It was selected in 2 of 5 runs.
- Its own domain was cited in 1 of 4 source-visible runs.
- A declared competitor appeared above it in 3 of 5 ordered answers.
- The answer format changed in 2 of 5 runs.
That is more useful than saying the brand “ranked well.” It separates presence, recommendation, citations, competitive placement, and volatility.
When the baseline is unclear, choose the remedy that matches the failure:
| Measurement problem | Better next step | Why |
|---|---|---|
| One important prompt produces mixed answers | Add runs under the same condition | The uncertainty is inside that prompt-platform row |
| The topic is represented by too few buyer intents | Add prompts in missing groups | Repeating one prompt cannot repair weak topic coverage |
| The panel is mostly branded | Add unbranded discovery and recommendation prompts | Recognition is not discovery |
| Competitors rotate across runs | Add runs and review whether the prompt is too broad | The shortlist may be unstable |
| Sources rotate while the answer claim stays similar | Track answer framing and citations separately | Claim stability and source stability are different signals |
| A launch or material change needs faster detection | Increase collection cadence temporarily | The issue is timing, not prompt coverage |
Red flag: rerunning until the brand appears and saving only that answer is selection bias, not baseline measurement.
Open the First Trend Window Without Moving the Setup
Cycle 0 records the starting level and the noise around it. Cycle 1 is the first directional comparison, but only if it repeats the same measurement contract after the baseline has been archived. The broader framework for measuring AI visibility as a trend applies only after comparable run sets exist.
Keep these fields comparable:
| Comparison control | Cycle 0 and Cycle 1 requirement |
|---|---|
| Prompts | Same wording and versions |
| Platforms and modes | Same answer surfaces, source conditions, and session rules |
| Market and language | Same localization context |
| Competitors | Same declared set; new observed brands remain separate |
| Labels | Same classification definitions and review rules |
| Run count | Like-for-like run sets for each compared condition |
| Denominators | Same eligibility rules for each metric |
| Observation span | Comparable window length and collection pattern |
| Evidence | Raw answers or excerpts preserved for both cycles |
Record the optimization date and the exact work performed between cycles. Also annotate known platform, model, collection, or scoring changes. Those notes help diagnosis, but timing alone does not prove that the optimization caused the movement.
An illustrative comparison might look like this:
| Signal | Cycle 0 | Cycle 1 | Cautious interpretation |
|---|---|---|---|
| Presence | 2 of 5 runs | 4 of 5 runs | Directional improvement in presence |
| Recommendation | 1 of 5 runs | 1 of 5 runs | More mentions did not improve selection |
| Own-domain citations | 0 of 4 source-visible runs | 2 of 4 source-visible runs | Source visibility may be improving |
| Competitor above brand | 4 of 5 ordered answers | 3 of 5 ordered answers | Competitive pressure remains |
| Answer volatility | High | Medium | The condition is becoming easier to monitor |
Two comparable cycles can show a directional difference. They do not establish a durable trend or an optimization effect. Require recurrence across later comparable cycles and check whether related signals support the same story. A presence increase with flat recommendation status is not the same as improved buyer consideration.
If a prompt, mode, competitor set, label, denominator, run count, or observation span changes, version the setup. Do not stitch the old and new data into one clean line.
Decision rule: Cycle 1 is the first comparison window, not proof. Report direction, distribution, affected segment, and uncertainty before assigning causation.
Use the Baseline as an Action Gate
The baseline is complete when it can route a finding to the right next step. It should not automatically generate content work.
| Baseline pattern | Better action | Why |
|---|---|---|
| Stable absence from core discovery prompts while declared competitors recur | Inspect category fit, competitor framing, and visible source evidence; optimize only when the gap maps to a controllable layer | The repeated in-scope pattern may represent a real discovery gap |
| Brand presence rotates across identical runs | Rerun, segment, or narrow the prompt | Volatility is not a clean content brief |
| Brand is mentioned but a competitor is repeatedly selected | Review the stated recommendation rationale and comparison evidence | The issue is consideration, not simple presence |
| The answer contains outdated or misleading product facts | Run an accuracy audit before visibility optimization | A factual risk should not be treated as a ranking problem |
| An owned page is cited but the answer remains vague | Inspect whether that page states category, use case, proof, and limitations clearly | The evidence layer exists, but the claim may be weak |
| Third-party sources repeatedly support competitors and omit the brand | Inspect the recurring source set before rewriting unrelated owned pages | The gap may sit outside the website |
| The brand is absent from an adjacent or out-of-scope prompt | Reclassify, monitor, or exclude the prompt | Not every absence deserves optimization |
| Platforms disagree while each platform is internally stable | Keep the platform segments separate and choose platform-specific actions | A blended average would hide the real difference |
Use this action sequence for every candidate optimization:
- Confirm comparability. Check prompt version, surface, mode, market, language, competitor set, labels, run count, and denominator.
- Confirm scope. Decide whether the prompt is core, adjacent, or out of scope.
- Inspect raw evidence. Read the answer and the excerpt that created the label.
- Separate the signal. Identify whether the issue is presence, recommendation, position, citation, accuracy, sentiment, or volatility.
- Check recurrence. Decide whether the pattern is isolated, mixed, or stable enough for the risk of the proposed action.
- Choose the route. Optimize, inspect sources, audit accuracy, rerun, monitor, exclude, or rebuild the setup.
The stronger the action, the stronger the evidence should be. One captured factual error can justify immediate accuracy review. A major content rewrite, competitive positioning change, or source campaign should require a recurring pattern under stable conditions.
Red Flags That Make the Baseline Unusable
Some setups should remain exploratory. Do not formalize a baseline when the category is disputed, the buyer questions are still being discovered, the competitor set cannot be agreed, or the team cannot repeat the same platform conditions after optimization.
| Red flag | Why it breaks the baseline | Better decision |
|---|---|---|
| Prompts change during Cycle 0 | The system is answering different questions | Version the panel and restart the cycle |
| Branded and unbranded prompts share one score | Recognition hides discovery weakness | Report the groups separately |
| ChatGPT, Gemini, Perplexity, and Google surfaces are averaged first | Platform-specific behavior disappears | Review each surface before any summary |
| Source-visible and no-source answers share a citation denominator | Citation coverage becomes meaningless | Limit citation metrics to eligible rows |
| Competitors are added after results appear | Share and gap calculations use a moving base | Keep new names as observed competitors until the next version |
| Screenshots are the only archive | Labels, modes, denominators, and prompt versions cannot be audited | Store structured row-level evidence |
| The best run is reported instead of the run set | Selection bias replaces measurement | Show x-of-n distributions and outliers |
| Every appearance is called a rank or recommendation | Different answer formats are forced into one label | Separate presence, prominence, order, and selection |
| A composite score has no drilldown | Movement cannot be diagnosed | Expose component metrics and evidence |
| Cycle 0 and Cycle 1 use different windows or run counts | Apparent movement may come from the setup | Recollect a comparable set or start a new line |
| A post-change increase is presented as causal proof | Timing is mistaken for attribution | Report directional movement and seek recurrence |
If several of these red flags are present, do not optimize from the report. Rebuild the measurement contract first. More data collected under inconsistent rules will create more precision-looking noise, not a stronger baseline.
Practical Takeaway
A reliable AI visibility baseline begins before optimization. Freeze the scope, organize prompts by buyer intent, declare competitors, separate platforms and modes, capture auditable rows, and measure starting noise. Use one run as a snapshot, three as a quick volatility check, five as a cautious operational read for important prompts, and ten or more for volatile or high-stakes conditions.
Then protect the first comparison window. Cycle 1 must repeat the Cycle 0 setup closely enough to show directional movement without hiding changes in prompts, platforms, competitors, labels, denominators, run counts, or observation span. Act on recurring, in-scope patterns with a controllable next step. Monitor, rerun, or rebuild the baseline when the evidence cannot yet support the decision.
