Chapter II — Evidence

The Boundary of a Claim

Evidence does not make a claim larger; it sets its limits. Seeing one answer is not the same as knowing a system's behaviour. A screenshot records a moment, not a general disposition. A brand may be mentioned in an answer, a text may be cited, a business may be recommended, or a customer may say that AI influenced a decision. Each is an observation. None, by itself, warrants a general conclusion.

“We appeared”, “we were recommended”, “we outperformed our competitors” and “GEO increased our sales” are not statements of equal weight. Each adds a new consequence to the one before it, and each new consequence requires a broader, more qualified body of evidence. A single model response can establish only that the response occurred. A customer's account may establish that AI played a part in the decision, but not that the system alone produced the sale. A rise in traffic may establish that more visits occurred during a given period, but not that a particular intervention caused the increase.

A claim may travel only as far as its evidence can carry it.

1. Evidence Is Not Decoration; It Is a Load-Bearing Relationship

No record is inherently strong. Its strength depends on the sentence it is being asked to support. A single screenshot that preserves the date, prompt and complete response may support the claim: “In its response to this prompt on this date, this system used the brand's name.” The same record cannot support: “Generative AI systems strongly recommend our brand.” The observation has not changed; the claim has expanded.

Data need not be altered to turn a genuine observation into an invalid conclusion. Erasing the boundary is enough. Evidence discipline therefore judges the sentence before it judges the data. “Our name appeared” is an observation; “we were represented accurately” concerns representation; “we were recommended more often than the selected competitors” is a comparison; “our intervention caused the change” is a causal claim; and “this representation brought us profitable customers” is a claim of commercial impact. One record cannot carry all five.

If the evidence cannot bear the load, the sentence is reduced. Uncertainty is not concealed, and missing ground is not covered with the language of success. A claim that exceeds its evidence is not polished. It is rejected.

This is not anxious overcaution. Early research into generative search systems showed that the presence of sources in an answer does not guarantee that those sources support every claim the answer makes.[2] A leading provider's own documentation likewise acknowledges that generative answers may contain inaccurate information, fabricated citations or unwarranted certainty.[6] Findings about particular systems at a particular time cannot be generalised to every product in use today. The distinction they reveal, however, endures: the presence of a citation and the adequacy of its support are not the same thing.

2. Three Domains of Evidence

GEO makes statements about three distinct realities: the world, generative representation and commercial outcomes. These domains may be connected, but they are not interchangeable.

2.1. Evidence about the real world

Evidence about the real world carries the underlying truth about a person or organisation. Does the company genuinely hold the certification? Is the hotel in fact intended for adults? Does the product provide the stated feature? Can the business serve the territories it claims? The answer to such a question does not come from a generative system. A system may repeat a claim; repetition is not verification.

The appropriate evidence about the real world depends on the claim. Records from competent authorities, valid certificates, contracts, auditable performance data, official corporate records and reliable independent publications may all carry weight. A business's own page is also a record, but it does not turn the business's claim of superiority into an independent fact. The source must still be assessed for independence, currency, scope and direct relevance to the claim.

2.2. Evidence of representation

Evidence of representation shows what a system produced under specified conditions: which system and interface, which prompt, which date, which language, which session conditions, which complete response and which sources? It does not show what the world is like. It shows how the system reconstructed the world.

A company's possession of a certificate does not prove that the system knows about it accurately. The existence of an official service page does not show that the system associates the business with that service. Correct information may have been published while the representation remains wrong. Representational integrity cannot be judged until evidence about the real world is compared with evidence of representation.

2.3. Evidence of outcomes

Evidence of outcomes traces what representation did at the level of the user and the business. Did the user see the answer? Did it influence the decision? Did the user contact the business? Was the enquiry qualified? Did a sale take place? Was delivery profitable and successful? Did the relationship recur? Each step requires its own record.

An invoice proves that a transaction took place, not that AI created the sale. “I found you on ChatGPT” is a signal of influence; it does not eliminate every other point of contact. A referral record may show a movement of traffic, not the whole decision journey. Outcome evidence often establishes only a partial relationship between events. A partial relationship is not complete causation.

Table 2.1 — The function and boundary of the three evidence domains

The right document, used in the wrong evidence domain, produces an invalid conclusion.

3. From Signal to Claim

“We have data” says nothing about the status of that data. The Framework distinguishes five stages: signal, observation, anecdote, evidence and claim.

A signal is a trace worth investigating: the brand appears in an answer, traffic rises, a customer mentions an AI system, or a competitor is absent from certain responses. A signal creates a question. It does not settle one.

An observation is a signal whose context has been recorded. The system, prompt, date, language, complete output and observer are known. It establishes what happened, not a general pattern of behaviour.

An anecdote turns the limits of a single event into a broader story. The event may be genuine while the generalisation remains unsupported. One customer's mention of AI does not establish that “AI brings us customers”; one recommendation in one response does not establish that “the system recommends us”. An anecdote need not be false. It is simply insufficient to serve as a rule of behaviour.

Evidence is a body of records arranged to support a defined claim. Its raw material is available, its context preserved, its comparisons made, its contrary results retained and its limits stated. Evidence does not remove uncertainty. It shows where uncertainty begins.

A claim is the precise, testable sentence the evidence is asked to carry. “GEO works” may be a slogan, but it is not a testable claim. “Within the defined set of thirty prompts, accurate attribute representation increased between these two measurement periods” reveals its universe, comparison and boundary.

4. Mention, Citation and Recommendation

These three observations are often presented as rungs on a single ladder of success. They are not the same event, and they do not carry the same evidential weight.

A mention occurs when an entity's name appears in an answer. The system may name the entity in an accurate or inaccurate context, confuse it with another entity, or use it as a negative example. A name's presence establishes linguistic presence, not understanding.

A citation occurs when the system associates a particular text with an answer. That association requires two separate tests. The citation test asks which claim the source has been attached to; the support test asks whether the source actually carries that claim. A source may be sound yet fail to support the sentence. It may concern another entity, be outdated or have too narrow a scope. To be cited is not necessarily to be supported by the citation.

A recommendation occurs when the system presents an entity as suitable for the user among a set of options. It is closer to a commercial outcome, but it is not one. It arose within a particular prompt, system, language, time and field of alternatives. A recommendation is evidence of decision influence, not evidence of purchase.

The boundary between the stages is clear: a mention does not by itself prove accurate representation; accurate representation does not prove recommendation; recommendation does not prove contact; contact does not prove a sale; a sale does not prove profitability; and profitability does not prove sustainability.

5. Five Conditions of Evidence

Five conditions must be considered together before an observation can support a public claim.

5.1. Context

Every observation records the system and mode of access, the date and time, the complete prompt, the language, the complete output, the sources shown and the observer. Where known, it also records conditions that may materially affect the result, such as account type, whether the conversation was new or continuing, and whether memory or search was active. If the system is unknown, the record cannot be compared; if the prompt is unknown, the observation cannot be retested; if the date is unknown, change cannot be followed. Output without context is not evidence. It is a memory.

5.2. Traceability

Evidence must be open to independent examination. The complete prompt and output are preserved; source links, access dates and, where necessary, archive copies are recorded; raw data is separated from human interpretation; and a synthetic example is never presented as a real observation. A cropped screenshot usually reveals only the area its narrator chose to show. A record that cannot be traced cannot be audited. A record that cannot be audited cannot support a consequential public claim.

5.3. Behavioural reproducibility

In generative systems, reproducibility does not mean seeing the same words every time. It means seeing the same representational behaviour recur under comparable conditions. The sentence may change while the wrong category persists; the wording may change while the same material limitation is omitted. A single record is an event. Predefined, controlled repetitions may establish a behavioural pattern whose scope is stated. What must recur is not the sentence, but the behaviour of representation.

5.4. Comparison and counter-evidence

A test that seeks only favourable results does not produce validation; it produces selection. Under the same conditions, the brand should be compared with relevant alternatives. The test set should also contain neutral prompts that do not name the brand, contrary prompts in which a competitor or another category may be more suitable, and exclusion prompts in which the brand should not appear. If contrary results narrow the claim, that is not a weakness in the method. It is evidence discipline doing its work.

5.5. Boundary

Every record must answer explicitly: “What does this observation not prove?” Behaviour in one system does not prove behaviour in others. A result in one language does not prove the same result in another. A week of observation does not prove permanence. A recommendation does not prove customer acquisition. A customer's account does not prove full causation. A sale does not prove sustainability. A record that cannot say what it does not prove does not yet understand what it does prove.

6. Claim Thresholds

The evidence threshold rises as a claim grows. The Framework uses five thresholds:

1. Observation claim: “This system used the brand's name in response to this prompt on this date.” The complete prompt, date and raw output may be sufficient.
2. Behavioural-pattern claim: “Across the defined prompt set, the brand was usually represented in the correct category.” This requires a controlled set, repetitions, a time range, coding criteria and records of deviation.
3. Comparative claim: “In the same test set, the brand was represented with accurate attributes more often than the specified alternatives.” Every party must pass through the same system, time, prompts and evaluation protocol.
4. Intervention or causal claim: “The intervention contributed to the change.” This requires before-and-after records, an intervention log, an adequate observation window and consideration of alternative explanations. To happen after something is not to happen because of it.
5. Commercial-impact claim: “During the defined period, AI-mediated discovery was associated with qualified enquiries and sales.” Customer accounts or referrals, CRM matching, enquiry quality, sales, profitability and other contact channels must be examined together.

These thresholds are not a scoring ladder. Reaching a higher threshold does not mean that better work was done; it means that a broader sentence is bearing a heavier load. If the evidence reaches only the first threshold, the correct conclusion is to stop there.

7. Four Fields of Counter-Evidence

A sound test universe contains four fields. The supportive field comprises prompts in which the entity should properly appear. The neutral field explores need and category without naming the brand. The contrary field covers conditions in which another option may be more suitable. The exclusion field covers contexts in which the entity should not appear at all.

The exclusion field is particularly important. Appearing everywhere is not success; appearing in irrelevant contexts is representational spillover. The Framework measures not only when an entity is selected, but also when it withdraws. Work that does not seek counter-evidence is not testing reality. It is confirming itself.

8. Teaching Case: Aurelis Packaging

Teaching simulation — synthetic data. This case does not represent a real business, source page or system response.

Aurelis Packaging is a fictional manufacturer of industrial packaging. A generative answer describes the company as “carbon-neutral, an industry leader and a user of 100 per cent recycled material”, citing the company's sustainability page as its source.

Table 2.2 — Synthetic support test for Aurelis Packaging

The citation is genuine; most of the conclusion is not carried by the source. The system's expansion is not the only risk. If the business repeats this favourable account in its own campaign, it will be treating generative output as evidence about the real world. The Framework does not approve the claim because a source exists. It returns every sentence to the limit the source can support.

The strongest publishable statement from this case is: “Aurelis Packaging reports that it is running a pilot at two facilities, using 40 per cent recycled material in one specified product line.” The claims of carbon neutrality and leadership are rejected because no independent evidence of suitable scope is available.

9. The Decision of Evidence

Every material claim is assigned, through the Evidence Record described in Appendix C, one of four decisions: rejected, further observation required, permitted only as an internal hypothesis, or publishable in limited form. The record is not completed in order to approve a claim. It is completed to decide whether the claim may be made.

Language is integral to that decision. Instead of “AI sees us as the leader”, write: “In three specified prompts, the system described the brand as a leader; no independent support was found.” Instead of “GEO increased sales”, write: “An increase in sales was observed after the intervention; causation was not established.” Limited language may look less spectacular. It carries more weight.

A record that fails the evidential threshold need not be discarded. A single output, a cropped screenshot, an unexplained score, a rise in traffic or a customer's account may be a signal that begins an investigation. What it must not be given is a burden it cannot carry.

The judgment of this chapter is:

Without evidence, there is no claim. With weak evidence, there is only a narrow claim. A strong claim can be made only on strong, traceable evidence whose limits have been stated.

The Centre established what is being represented. Evidence has now set the limits of what may be said about that representation. The next task is to define the conditions, units and comparisons by which the observations carrying this load are to be recorded.