Sentiment Delta

Sentiment Delta #

Machine Relations defines Sentiment Delta as an observation metric. It compares a brand's declared positioning with descriptions returned by AI systems under a documented test. The metric records differences in category, attributes, and framing. It does not infer an engine's private reasoning, source weights, or effect on buyers.

Machine Relations definition: An observation records an engine response to a declared query at a recorded point in time. Comparable collections keep the query set, collection procedure, brand reference statement, and coding rules fixed while recording the engine, model or product surface, date, and response.

What the Research Establishes #

A large-scale empirical study comparing Google Search with responses from GPT-4o, Claude, and Perplexity found differences in consulted source domains, source types, query intent, and freshness (Navigating the Shift, 2025).

Premise: this page does not convert that bounded result into a universal source-mix claim for brand queries.

A separate study analyzed 55,936 queries across six LLM-based search engines and two traditional search engines and reported that 37% of the observed domains were unique to the LLM-based systems in that sample (Zhang et al., 2025).

Premise: Machine Relations measures each declared surface directly rather than treating a search result as a proxy for all AI systems.

Research on cognitive biases in LLM-based product recommendation found that changes to product-description language affected recommendation rates and rankings in the tested models; the reported effects varied by framing and model (Filandrianos et al., 2025).

Premise: that experiment is a bounded recommendation test. It is not evidence that third-party coverage determines brand descriptions or business outcomes.

Measurement Protocol #

A reproducible Sentiment Delta observation records:

1. Brand reference statement — a versioned set of categories and attributes the brand declares for itself.
2. Query set — the exact category-relevant prompts used for collection.
3. Collection context — engine, model or product surface, date, locale, and any retrieval setting available to the researcher.
4. Raw responses — the unedited descriptions returned for each query.
5. Coding result — the declared category, attributes, and framing present or absent in each response.

Machine Relations doctrine: report the component differences before creating any aggregate score. An aggregate can hide whether a change came from category, attribute, omission, or framing. If an aggregate is used, its formula and missing-data treatment must be published with the result.

Delta Types #

The following types are Machine Relations coding definitions, not claims about engine internals:

Type	Recorded difference	Arithmetic illustration
Categorical delta	Returned category differs from the declared category	Reference: "PR agency"; response: "marketing firm"
Attribute delta	A declared attribute is absent or a different attribute appears	Reference emphasizes speed; response emphasizes price
Framing delta	Emphasis or evaluative language differs	Reference says "enterprise"; response says "for small teams"
Omission	The response contains no description of the brand	Brand is absent from the collected answer

Illustration: these examples show the coding rule only. They are not observed benchmarks or a ranking of consequences.

Reporting Sentiment Delta #

For each collection window, report:

• the number of eligible responses;
• category matches and mismatches;
• declared attributes present, absent, or replaced;
• framing differences under the published codebook;
• omissions and unreadable or unavailable responses; and
• uncertainty or coder disagreement.

Premise: a response with no recorded difference has zero observed delta under that test. It does not prove that every possible query, engine, or user receives the same description. A nonzero delta records a difference without diagnosing its cause.

Comparing Windows #

Premise: window comparisons use the same query set, reference statement, coding rules, and collection settings where possible. Unavoidable changes, including model or product updates, remain part of the record rather than being attributed to the brand.

Machine Relations premise: source changes, model changes, retrieval changes, prompt changes, and ordinary sampling variation are competing explanations for a changed result. Sentiment Delta alone cannot separate them. A causal claim requires a separate design with controls appropriate to that claim.

Sentiment Delta and Share of Citation #

Premise: Share of Citation and Sentiment Delta answer different measurement questions. Share of Citation records appearance frequency under its own declared protocol. Sentiment Delta codes differences in descriptions when a usable description is present. Report them separately; neither metric establishes the cause or business effect of the other.

---

FAQ #

Can Sentiment Delta be positive or negative?

Only if the published codebook defines a direction. The default MR protocol reports the type and magnitude of observed differences without treating alignment as inherently favorable or unfavorable.

How long does Sentiment Delta take to change?

Premise: the metric reports results from declared collection windows and makes no promise about when an engine will return a different description.

Is Sentiment Delta the same across all AI systems?

Do not assume so. Measure each declared system and collection context separately, then compare only under the published protocol.