Does ChatGPT Recommend My Business? How to Measure It

One missing recommendation is a clue, not a verdict. Owners who search does ChatGPT recommend my business often see one competitor-filled answer and assume the result describes a settled market position. It describes only one generated output. A useful conclusion needs repeated buyer questions, captured evidence and a distinction between a name appearing and a business being recommended.

Quick Answer. The answer to does ChatGPT recommend my business comes from a repeatable sample, not one prompt. Record several real buyer questions, repeat each run, and compare mentions, recommendation context, competitors and cited sources. One absence is a snapshot; a recurring pattern under the same protocol is evidence you can investigate.

Last updated: 27 August 2026

A missing answer is evidence, not a verdict

An answer that omits the business tells you exactly one thing: that output did not include it. The result may still matter. It can reveal a question where competitors repeatedly appear, a mismatch between the buyer's wording and the business's evidence, or an assistant that returns no usable recommendation. What it cannot reveal alone is a stable level of visibility.

Generated answers are assembled, not retrieved as a fixed list. The foundational GEO research describes generative engines as systems that gather and synthesise information from multiple sources ¹. Each output is therefore a fresh response to a particular query and context. It is not a stored league table that a business occupies until the next update.

The practical response has two branches. If the name appears consistently in a relevant, favourable context, preserve the evidence and check whether the description is accurate. If the name is absent or appears inconsistently, expand the sample before diagnosing. Both branches require the same discipline: save what the assistant said, not what you remember it saying.

A small specialist consultancy may appear for “best data migration adviser for a regulated manufacturer” but not for “best IT consultancy”. Those questions describe different buyers and different evidence needs. Combining them into a single present-or-absent verdict would erase the very distinction the owner needs to see.

A mention and a recommendation are different events

A mention is the business name appearing anywhere in the answer. A recommendation is narrower: the answer connects that business to the buyer's need and presents it as an option worth considering. The difference matters because a name can appear as a comparison point, a source, a warning, an unrelated namesake or a provider that the answer explicitly rules out.

Use three labels when coding an answer:

• Recommended — the business is presented as suitable for the stated need.
• Mentioned only — the name appears, but the answer does not endorse its fit.
• Absent — the business does not appear in the usable answer.

Accuracy belongs in another field. A recommendation that assigns the wrong location, speciality or service is visible but harmful. A clean measurement therefore records “recommended” and “accurate” independently. Collapsing them into one score can make a confidently wrong description look like success.

The wording around competitors also deserves capture. An answer may group several businesses without ordering them, explain why each fits a different buyer, or describe one as the conventional choice and another as a specialist. Those are distinct recommendation contexts, even when the same names appear.

For a deeper account of the mechanisms behind those differences, see why an AI recommends a competitor instead. That explainer separates relevance, prominence, corroboration and freshness rather than treating every absence as the same problem.

Decide whether you saw a snapshot, a pattern or bad data

The first decision is not “how do I improve the website?” It is “what kind of evidence do I have?” A screenshot from one prompt is a snapshot. Repeated outcomes across a held-constant protocol form a pattern. A wrong namesake, truncated answer or mismatched location is a data-quality problem and should not be counted as an ordinary absence.

In prose, the flow is: Usable answer? The NO branch leads to Record no usable answer. The YES branch asks Business named? An absent name leads to Repeat the protocol. A present name leads to Relevant recommendation? The NO branch means Code mention only; the YES branch means Verify description and sources. The repeated sample supports Snapshot or pattern, followed by Repeat, compare, diagnose.

This order prevents two common mistakes. First, a refusal or cut-off is not silently converted into “brand absent”. Second, a name appearing in an irrelevant paragraph is not promoted to “recommended”. The rows remain comparable because every run passes through the same decisions.

The table is deliberately conservative. It does not minimise the business importance of a recurring absence. It makes the finding defensible, which is more useful than a dramatic claim that cannot survive a rerun.

The names can change between runs

Variation is expected enough to measure, but it should not be used as a universal excuse. The controlled evidence shows specific kinds of movement in specific studies. A 2026 preprint audited 2,000 runs across ten personas, eight prompts and three model configurations with repeated samples. Adding persona context reduced recommendation-set similarity by 0.12–0.20 Jaccard points against the study's same-persona baseline ².

That result does not say every business will move by those amounts. It shows why a buyer persona is part of the protocol rather than colour added to a prompt. A solo founder, a procurement lead and a local consumer can value different constraints while using similar category words.

A separate seven-day audit examined 48 authentic queries across four topics in ChatGPT, Bing Chat and Perplexity. It found query- and topic-sensitive sentiment, alongside commercial and geographic source bias in that sample ⁴. Again, the scope matters: three named systems, four topics and seven days. The study supports repeated, context-aware measurement; it does not provide a current visibility rate for a local business.

Location, language and wording should therefore be changed deliberately, never casually. If a protocol is meant to test one buyer scenario, hold them constant across repeats. If the business wants to compare scenarios, create separate groups and label them. Mixing several variables inside one group produces movement that cannot be interpreted.

Evidence makes a business eligible to be named

An assistant needs material that supports a relevant answer. The material can include a clear service page, precise location and eligibility information, consistent business facts, third-party coverage, documented expertise and recent evidence. These are not switches that force inclusion. They are inputs that make a recommendation easier to justify.

Generative engine optimisation is the practice of improving visibility within synthesised answers. The plain definition of generative engine optimisation explains where it overlaps with search optimisation and where it does not. The foundational GEO paper introduced a visibility framework and tested content strategies on GEO-bench; its reported benchmark effect varied by domain ¹. That is an experiment, not a promise about ChatGPT or a forecast for one business.

The useful distinction is between owned and earned evidence. Owned pages explain what the business does, for whom, where and under what constraints. Earned sources provide independent corroboration. Both can be weak: an owned claim may be vague, while a third-party listing may be stale or describe the wrong service.

Verifiability must be checked, not assumed from polished citations. A 2023 human audit of Bing Chat, NeevaAI, Perplexity and YouChat found that 51.5% of generated sentences were fully supported by citations and 74.5% of citations supported the associated sentence ³. Those figures describe the audited 2023 systems, not current ChatGPT. Their operational lesson is still direct: inspect whether a citation covers the claim and whether the source actually supports it.

Measure the pattern without overstating it

A sound protocol begins with real buyer questions. These are questions a prospect would ask before choosing, not vanity prompts that insert the business name and invite recognition. Include different stages of the decision: category discovery, suitability, location, constraints, alternatives and specialist needs.

The next step is repetition. A single run cannot reveal whether an outcome recurs. The protocol should specify the assistant, model or product surface where visible, language, location, persona wording, account state if relevant, date, and number of repetitions. No sample size turns the output into a permanent rank; repetition simply makes the observed pattern less dependent on one generation.

The full capture row should contain enough information for another person to understand and recheck it:

For the operational sequence, use the repeatable method for checking whether AI mentions a brand. It covers question design, repeated runs, capture, scoring and rechecks without converting a manual observation into an audit claim.

The outcome should be reported as a fraction within the protocol: for example, “named in seven of fifteen usable runs for these three buyer questions during this test”. A sentence like “ranked second in AI” invents an ordering and permanence the evidence does not contain. If the assistant gave an ordered list, record that order as part of the answer, not as a platform-wide rank.

Fix the narrowest measured gap first

The first improvement should follow from a recurring, interpretable gap. If the business is absent only for one specialist use case and its site never explains that service, the content gap is specific. If it is described with an old address across several answers, entity correction comes first. If competitors are supported by independent sources while the business has only self-published claims, more owned copy alone may not solve the evidence gap.

A bounded change is easier to evaluate than a general rewrite. Add one service page, correct one set of conflicting facts, publish one evidence-backed case example, or earn one relevant independent reference. Record the change and its publication date. Then rerun the same questions with the same settings after the evidence has had a reasonable chance to be available.

The comparison should use the original capture fields. Look for a movement in mention frequency, description accuracy, competitor set or source use. Even a favourable change does not prove a single cause, because the systems and available web evidence may also have changed. It does tell you whether the measured output moved after the intervention.

Some gaps require no content action. A mixed result across repetitions may call for a larger sample. A wrong namesake requires identity disambiguation. A finding confined to one unusual vanity prompt may deserve no work at all. Measurement protects the team from spending effort on the most emotionally vivid screenshot rather than the most consistent buyer-facing gap.

Avoid promises the evidence cannot support

No ethical visibility plan guarantees inclusion, traffic or revenue. The assistants are changing systems, the available sources change, and the buyer's question changes what counts as relevant. A well-supported page can become a better candidate for synthesis without becoming entitled to a place in every answer.

Avoid manipulative tactics such as fabricated authority, invented reviews, copied comparison pages or claims written only to trigger a model. They weaken the evidence a reader needs and create reputational risk. The business should improve the truth available about its real offer: precise scope, consistent facts, useful explanations and genuine third-party support.

The same caution applies to competitors. A recurring competitor mention is diagnostic evidence, not proof that the competitor is objectively better. It may reflect stronger relevance to the prompt, more prominent evidence, fresher corroboration or a contingent incumbent effect. The correct next move is to inspect the evidence path, not to imitate the competitor's language blindly.

Finally, separate visibility from business value. A business can be named often for the wrong service, wrong location or wrong reasons. The best result is an accurate recommendation for a real buyer need, supported by evidence the buyer can inspect. That is a narrower and more defensible target than “win AI”.

Move from one query to a measured audit

Begin with the question that caused concern, then widen it into a buyer-question set. Separate personas, repeat the runs, capture the complete answers and code recommendation context independently from accuracy. The result will show whether the original screenshot was isolated, recurring or unusable.

A fuller measurement should widen question coverage, assistant coverage and repetition coverage without pretending uncertainty disappears. The supplied full protocol uses twenty researched buyer questions, three assistants and five runs per question; the free check uses one question on one assistant twice and remains a snapshot. Its defensible advantage is the breadth of those three sampled dimensions. Every returned answer remains an observation under the named protocol rather than evidence of a durable rank.

When a business needs that broader scope, use the 20-question, three-assistant, five-run AI Answer Audit measurement. That scope defines how answers are sampled; it does not promise any further product analysis or a durable rank.