Every business that buys search or AI visibility work eventually asks the same question: can I trust what I’m being told? Not because the industry is full of liars — most people in it believe what they’re selling — but because almost nothing in it is labelled. Verified results, documented sightings, professional judgement and pure guesswork all arrive in the same confident tone, on the same slide, at the same day rate.
A news week in June 2026 makes this the right moment to draw the labels properly. Google published new Search Central guidance on third-party SEO advice and tools — including the blunt line “Google doesn’t evaluate third-party services” — and told businesses to check whether advice is grounded in evidence, anchored to official documentation, or merely implied to carry Google’s blessing. Read one way, it’s Google asserting authority over SEO, AEO and GEO advice. The guidance is useful for a different reason: it helps draw a boundary between evidence, observation, interpretation and speculation. What follows is that boundary, drawn in full.
The four levels of any SEO claim
Every claim you will ever hear from an SEO, AEO, GEO or AI visibility provider sits on one of four rungs. The industry’s habit is to blur the rungs. The entire skill of buying well is keeping them separate.
1. What they know — Evidence. Verifiable in your own first-party data. Rankings moved. Organic traffic rose. Enquiries came in. You can see it in Google Search Console, your analytics, your inbox. Nobody has to interpret it for you, and nobody can fake it for long.
2. What they observe — Observation. Documented behaviour in the wild: an AI Overview cited your page for a specific query on a specific date; an assistant named your business in a recommendation; a competitor’s content was used as the substrate for an answer. Observations are real and valuable — but they are data points with conditions attached, not proofs. Ask whether they are dated, recorded, and honest about what else might explain them. Most providers have stories. Very few have records.
3. What they think — Interpretation. Professional judgement: pattern recognition across hundreds of sites, trend analysis, informed opinion built on years of evidence and observation. This is where genuine expertise lives, and it is legitimate — provided it is labelled as judgement and can be debated. An interpretation that cannot tolerate the question “what would change your mind?” has quietly become the fourth thing.
4. What they guess — Speculation. Causal certainty about Google’s internals: why a page ranked, which signal carried the weight, what a change will definitely produce. It cannot be proven by anyone outside Google, no matter how confidently it is presented. Speculation isn’t always wrong. It is always unprovable — and the moment it is sold as knowledge, it stops being expertise and becomes salesmanship.
The same test, in plain English: what does your provider know, what do they observe, what do they think, and what are they guessing — and do they tell you which is which?
The quality of an SEO proposal is often determined less by which rungs it contains, and more by whether the provider clearly separates the rungs. Interpretation and hypothesis are not the problem — a consultant with no interpretations has no expertise to sell. The problem is interpretation dressed as evidence, and speculation dressed as anything at all.
One consequence of this is counterintuitive: the higher rungs are not always the more trustworthy ones to hear. A provider who says “I don’t know why this worked, but here is the evidence that it did” is often more credible than one who confidently explains every outcome. Confident explanations of Google’s internals are cheap. Verifiable results and honest attribution are not.
What can actually be known
Here is the boundary, drawn plainly. It is worth keeping.
| The claim | Can it be known? | Rung |
|---|---|---|
| Did rankings improve? | Yes — first-party, verifiable | Knows |
| Did organic traffic increase? | Yes — first-party, verifiable | Knows |
| Did enquiries and revenue increase? | Yes — first-party, verifiable | Knows |
| Was the page cited in an AI answer? | Usually — observable, documentable, surface by surface | Observes |
| Did AI citations increase over time? | Usually — if someone is recording dated observations | Observes |
| Is this pattern likely to repeat? | Debatable — judgement, only as good as the records behind it | Thinks |
| Why exactly did Google rank the page? | No — only Google knows | Guesses |
| What weighting does Google apply to any signal? | No — only Google knows | Guesses |
| Will this change guarantee a result? | No — anyone who says otherwise has already failed the test | Guesses |
One sentence sits underneath the whole table: nobody outside Google has Google’s ranking data. Not the biggest agency, not the most expensive tool, not this consultancy. That single fact quietly disposes of guaranteed rankings, proprietary “visibility scores” presented as ground truth, and we-know-the-algorithm claims. What survives it is everything legitimate: documented experience, dated observations, trend analysis, statistical confidence, case studies, and informed opinion labelled as informed opinion. The legitimate work has never needed the certainty claim.
Five questions to ask any provider before you hire them
The questions follow from the table. You do not need to understand the technology to use them.
- Where does this advice come from? Is it labelled as judgement based on data and experience, or anchored to official documentation? Either is fine. Ask for the evidence or the link.
- Where does this data come from? Whose crawler, whose criteria, whose scale? If you are shown a score, ask who built the scoring system and what it actually measures.
- Is anything implying Google approval? Google states plainly that it does not evaluate third-party services. Any badge, partnership framing or “Google-approved” language suggesting otherwise is a red flag by Google’s own definition.
- Are predictions labelled as predictions? Ranking and traffic forecasts are the vendor’s own. They may be informed and useful. They are not guarantees, and anyone presenting them as guarantees has already failed the test.
- Which rung is each claim on? Ask them to separate what they know, what they’ve observed, what they think and what they’re guessing. The provider who can answer that comfortably — and shows you dated records for the middle rungs — is usually the provider you want.
If a provider cannot clearly tell you which claims they know, which they’ve observed, which they think, and which they’re guessing — stop the conversation there.
Keep the five questions. They will outlast this news cycle.
The Observation Test, applied to a real outcome
Drawing these lines is easy to say and harder to do, so here is what it looks like applied to a real result — one documented in this site’s Observed Outcomes Register. Four questions, asked the same way every time.
What happened? In March 2026, a criminal defence firm’s page about Operation Soteria was cited in a Google AI Overview, with the firm’s logo displayed in the citation panel.
What do we know? The citation happened, on a dated, screenshotted occasion. The page had a named author with verifiable credentials, an opening that answered the query directly, defined terms, attributed claims, and structured content throughout. The firm carries strong independent corroboration — Chambers, Legal 500, law society listings.
What don’t we know? Which of those factors mattered most. Whether the citation was caused by any single one of them. Whether the same structure would produce the same outcome in a different sector. How long the behaviour will persist as Google’s systems change.
What is the most honest conclusion? The outcome is consistent with the hypothesis that well-structured pages on well-corroborated entities get selected as AI answer substrates — and one observation is a data point, not a pattern. That is why a register exists at all: observations recorded the same way every time, confounders named, so hypotheses earn evidential weight or get refuted as entries accumulate. Either outcome is useful.
Notice what that framing costs: nothing. The result is still commercially impressive. The client still won. Honesty about uncertainty does not weaken a genuine result — it is only fatal to results that were never genuine.
The buying behaviour that follows from this is simple. Don’t only ask for case studies. Ask to see the observation log: the dated, structured record of what was seen, when, under what conditions, with the unknowns written down. If the only evidence available is a polished case study written after the outcome, you are seeing a narrative. An observation log shows what was believed before the outcome was known. The answer to that request tells you more than the case studies do.
An honest example of uncertainty
It would be convenient if every example supported the position, so here is one that doesn’t resolve neatly. Through May 2026 this site shipped a rapid sequence of changes to its own pages — content additions, schema changes, structural edits, sometimes several in the same week. Some pages improved afterwards. We can observe exactly which pages and when, because the rankings are tracked and the releases are dated. What we cannot do is prove which change caused which improvement, because multiple variables moved at once. The outcome is on the first rung. The attribution is stuck on the third — it is judgement, debatable, and recorded as such rather than upgraded to certainty because certainty would sell better.
That paragraph is the whole discipline in miniature.
Any provider can show you their wins. Ask them to show you where their attribution ran out.
What this means, by who you are
If you own a business or a website — understand it. Use the table and the five questions. You now have a working test for every proposal that lands in your inbox, and you did not need to learn what an algorithm is to apply it.
If you lead marketing — seize it. Run the test across current suppliers and shortlists; budgets are being spent right now on claims from the bottom rows of the table. And start reporting visibility and referrals as separate numbers — AI citations and branded mentions on one line, clicks and conversions on another. As AI-mediated search grows, the gap between those lines widens, and the marketer who can explain the gap owns the conversation.
If you do the work — SEO, AIO, development — implement it. Label the rungs explicitly in everything you publish and report: evidence from first-party data, observations with dates and conditions, judgement marked as judgement. Anchor recommendations to official documentation or clearly-labelled experience. Google’s guidance has made that distinction explicit; the practitioners it hurts are the ones who blurred it, and the ones it helps never did.
The boundary is the value
Everything on this site is offered under the same discipline: observations with the confounders named, judgement labelled as judgement, and no number presented as Google data — because it isn’t.
Uncertainty is not the opposite of expertise. Pretending uncertainty does not exist is.
Nobody outside Google has Google’s ranking data. The question is not who claims certainty. It is who is honest about uncertainty.
