White-Label AI SEO: What Agencies Should Check Before Reselling

When you white-label AI SEO, you put your logo on someone else's measurement. The client reads the number as your claim. If the formula is wrong, the client blames you, not the vendor.

Most white-label packages bundle three things:

• A visibility score, usually one number per client, sometimes per engine.
• Prompt data: a list of tracked questions with mentions, citations or sentiment.
• Recommendations: the fixes the report tells the client to make.

Each part needs a different check. The score needs a published formula. The prompt data needs a defined sample. The recommendations need evidence behind them.

The labels vary (white label GEO, white label AI visibility, AEO reporting), and the labels mostly describe the same work. The checks below apply whatever the package is called.

A metric you can't check against anything is a claim, not a measurement. Several common AI visibility metrics fall into that group.

No position data exists for Google AI

Google's Search Console report for generative AI features shows impressions only, with no clicks, CTR or position. Any "AI rank" in Google has nothing to be validated against.

Retrieval fires at the engine's discretion

ChatGPT decides per question whether to search the web. Published estimates of how often it does range from about 18% (Profound) to 80.5% (cloro), and they measure different things. A score that doesn't say whether search was on blends two different kinds of answer.

The engine rewrites the prompt

OpenAI's help page says ChatGPT search rewrites your query into one or more targeted queries before it searches. A prompt tracker measures the question it typed, not the queries that actually ran.

Surfaces are not interchangeable

Google says AI Overviews and AI Mode "may use different models and techniques". Engines also cite differently: Anthropic cites text spans, Perplexity cites whole sources. A single blended score averages units that don't share a denominator.

Cited is not the same as used

OpenAI's API returns the URLs it consulted separately from the ones it cited, and the consulted list is often longer. Citation counts undercount influence.

Every tracker is a sample

In Otterly's 2,000-blog experiment, Microsoft's own AI Performance report credited one site with 42,817 citations, while Otterly's prompt-based tracker logged 3,934. The windows may not match, but the gap shows that prompt tracking sees only what it asks.

A critical survey of 45 GEO studies (arXiv, July 2026) found no technique with a stable, cross-platform causal effect on discoverability. A report that credits a score rise to a specific tactic is making a claim nobody has measured.

Ask these questions before you sign. A vendor that can't answer them in writing is asking you to vouch for numbers it won't explain.

These defects are invisible from a dashboard. 99Visibility found one in its own accuracy formula in September 2026: it summed penalty points with no denominator. An audit of 16 answers with a 12.5% error rate scored 90, one of 68 answers with a 10.3% error rate scored 65, and one of 200 answers at 10% scored 0. A client who bought more prompts would have watched the score fall. The fix was a rate-based formula, a formula version stored on every score, and a written note that its constants are chosen, not derived. It also states that any single score, including its own, averages answers given with and without live search. That is the level of disclosure to expect from any vendor whose numbers you resell.

AI answers pass through four stages. Knowing which one a metric reflects tells you whether anyone can act on it.

• Parametric: what the model knows without searching, frozen at a training cutoff. Nobody changes this in a quarter.
• Access: whether a crawler can fetch and store the page. The client and their CDN control this.
• Retrieval: whether the page enters the answer for a given query. The engine controls this, partly.
• Generation: which retrieved pages get used and cited. The model controls this.

In the client's control:

• Per-bot robots.txt rules. OpenAI says each of its crawler settings is independent: GPTBot covers training, OAI-SearchBot covers ChatGPT search.
• Firewall and CDN rules that may block a crawler robots.txt allows.
• Server-side rendering of key facts, since most AI crawlers don't execute JavaScript.
• noindex, nosnippet, max-snippet and data-nosnippet, which keep content out of AI answers.
• Consistent facts (prices, product names, locations) stated plainly on fetchable pages.

Outside anyone's control: whether search fires, where a page sits in the model's context, and what the model says. That is why an agency should never guarantee AI citations, white-labeled or not.

Run these in order. The early ones are cheap and rule out most weak vendors.

1. Get the formula in writing. Components, weights, denominators and constants. Note which parts the vendor calls measured and which are its model.
2. Run a repeat test. Audit the same brand with the same prompts twice in one day. Large swings mean the score reflects sampling noise.
3. Run a scale test. Double the prompt set. If the error rate stays flat but the score moves sharply, the formula has no proper denominator.
4. Inspect each engine column. Confirm the model, the retrieval state and whether Google AI Overviews, AI Mode and Gemini are measured separately.
5. Check the access audit. It should parse robots.txt per bot and also request pages as each crawler, because a firewall can return 403 to a bot robots.txt allows. Cloudflare changed its AI crawler defaults on 15 September 2026 for new domains and free-tier customers, so a client can lose access without editing anything.
6. Audit the recommendation library. Compare every recommendation against the next section before it goes out under your name.
7. Decide what you will say. Write down which numbers you present as measurements and which as estimates.

For portfolio-scale needs, see what an agency managing 10+ brands should look for in a tool. For presenting results, the agency client reporting playbook covers format and cadence.

Many white-label report templates still carry advice that primary sources contradict. Each one is a claim a well-read client can catch.

• "Add schema or FAQ schema to get cited." Google says structured data isn't required for generative AI search. Ahrefs' matched test of 1,885 pages found AI Overviews citations fell 4.6% after schema was added. FAQ rich results were deprecated in May 2026. Don't tell clients to remove existing markup either: it simply isn't a lever for AI answers.
• "Publish an llms.txt." Google says Search ignores it, and no engine documents reading it.
• "Improve your E-E-A-T score." Google says E-E-A-T isn't a specific ranking factor. No score exists.
• "Build domain authority." Domain Authority is a Moz metric. No engine computes it.
• "Allow GPTBot so ChatGPT finds you." GPTBot governs training. ChatGPT search uses OAI-SearchBot.
• "Chunk content" or "write AI-friendly copy." Google rejects both and says there is no ideal page length.
• "GEO gets up to 40% more visibility." That figure came from a lab engine with sources already loaded into context.
• "Ping IndexNow for faster AI citations." The spec never mentions AI, and a 200 response only confirms receipt.

What to keep: rendering, per-bot crawler permission, network reachability, index eligibility, snippet controls, and correcting wrong facts. When engines cite other sites, being named on those sites has a strong association with appearing in answers, with no promised effect. For a fuller scope, see what AI SEO services should include.

• Making a blended score the headline KPI. It hides which engine moved. In Otterly's experiment, one engine cited a site thousands of times while four barely did.
• Reselling vendor case studies. Before-and-after charts rarely control for brand size or retrieval state. Read how to judge GEO case studies before reusing one.
• Comparing scores across formula versions. A trend line that spans a formula change shows the change, not the client.
• Treating a robots.txt allow as access. It's necessary, not sufficient. Pair it with a live request as each crawler.
• Benchmarking against much bigger brands. Brand prominence is the strongest association in the citation data and nobody can buy it quickly. Compare like-sized competitors.
• Letting the template write the narrative. Edit every sentence that states a cause. If it isn't an access fix or a measured audit result, soften or cut it.

Is white label GEO different from white label AI SEO? Mostly in name. Vendors use GEO, AEO, AI SEO and AI visibility for overlapping products: tracking how AI engines describe a brand and recommending fixes. Judge the package by its formulas, samples and recommendations, not the label. Two products with different names can share the same weaknesses, and two with the same name can differ completely in how they compute a score.

Can I promise clients AI citations if the vendor does? No. Nobody controls whether an engine searches, which pages it retrieves, or what the model writes. A vendor guarantee doesn't change that, and once it's under your brand the broken promise is yours. You can commit to fixing access problems, correcting wrong facts on the client's site, and reporting honestly what changed. Those are outcomes within your control.

How many prompts should a white-label report track? The evidence gives no correct number. What matters is that the set stays consistent between reports, the sample size is stated, and the formula gives the same score for the same error rate whether you track 20 prompts or 200. If adding prompts moves the score with no change in answer quality, the formula is the problem, not the client.

Should reports split Google AI Overviews, AI Mode and Gemini? Yes. Google says AI Overviews and AI Mode may use different models and techniques and return different links, and the Gemini app is a separate product again. A single "Google AI" column hides which surface changed. As of October 2026, ask the vendor exactly which Google surfaces it measures and how, since some have no official API.

What should I tell a client whose score dropped? Check the boring causes first. Did the formula version change? Did the prompt set or sample size change? Did search switch on or off for an engine? Then check access: a firewall rule, a CDN default such as Cloudflare's September 2026 change, or a stray noindex can remove a site silently. Report what you found, not a story about the algorithm.