Checking whether ChatGPT, DeepSeek or Doubao cites a brand requires more than running one branded query and saving a favourable screenshot. A reliable assessment uses a fixed set of buyer questions, repeats them under comparable conditions, and distinguishes between a passing mention, inclusion in a shortlist, a reasoned recommendation, factual accuracy and a traceable citation. The exact prompt, response, source links, date, market, language and competing brands should all be retained.
A mention, recommendation and citation are different signals
A brand name appearing in an answer is a mention. Inclusion alongside alternatives is shortlist visibility. A response that explains why the brand fits the user's stated needs is closer to a recommendation. A citation is separate: the answer provides a source that can be opened and that genuinely supports the relevant claim. Accuracy must also be scored independently because an answer may name the correct company while misstating its services, location, contact details or client work.
Step 1: build a query set around real buying needs
Do not test only a query such as “What is WEPR?” Buyers usually begin with a category, situation or risk. For each priority service, create brand, category, use-case, comparison and risk questions. Start with 20 to 30 commercially relevant questions. Keep roughly 70 percent stable so month-to-month results remain comparable, and refresh the remainder using sales conversations, new services and market changes.
- Brand: What services does WEPR provide?
- Category: Which GEO agencies operate in China?
- Use case: How should a Chinese brand improve AI search visibility overseas?
- Comparison: How does GEO differ from conventional SEO?
- Risk: What should a buyer avoid when selecting a GEO provider?
Step 2: control the test conditions
The same prompt can produce different answers by date, account, location, language, model, conversation context and web-search setting. Record the platform, exact prompt, date and time, language, login state, market, device and whether web search was enabled. Use a fresh conversation wherever practical. A brand appearing once and disappearing once is not enough evidence of a trend; repeat high-value queries and look for sustained movement.
Step 3: apply one scoring framework
- Presence: absent, incidental mention, shortlist or primary recommendation
- Accuracy: accurate, partly accurate, outdated, wrong or unverifiable
- Citation quality: none, irrelevant source, first-party source or credible independent source
- Competitive context: which alternatives appear and why
- Required action: improve facts, evidence, cases or conflicting information
Do not interpret answer order as a conventional search ranking. AI-generated responses do not necessarily follow the logic of a search results page. The stronger signal is whether the brand repeatedly enters the relevant consideration set with accurate, useful reasons.
Step 4: verify every source
Open each cited link. Check that it is accessible, directly supports the statement, remains current and comes from an identifiable source. Distinguish the company website, independent reporting, industry directories and aggregators. Watch for answers that recommend a company while citing only a generic industry article; that link does not validate the brand recommendation.
Step 5: report metrics with sample sizes
A basic visibility rate is the number of questions in which the brand appears divided by the total tested. Add shortlist rate, primary-recommendation rate, factual accuracy, valid-citation rate and competitor presence. Always keep the denominator. Segment results by service and buying stage; educational visibility without provider-selection visibility often signals a lack of cases, methodology, proof and procurement information.
What to do when the brand is absent
Do not respond by mass-producing near-duplicate articles. Confirm that the website clearly states the legal entity, services, expertise, methods, cases, contact details and update dates. Check evidence, internal links and crawler access, then fill the specific gap revealed by the query set. If an AI answer contains an error, correct conflicting facts across the website and authoritative profiles before retesting.
How often should monitoring run?
For many organisations, a monthly full test plus a weekly sample of high-value prompts is sufficient. Recheck sooner after a rebrand, service change, major publication or material factual error. WEPR separates brand presence, factual accuracy, source quality, competitive share and website enquiries instead of compressing everything into one score, showing whether buyers can find the brand, understand it correctly and continue their research.
Does one appearance mean the brand is indexed by AI?
No firm conclusion should be drawn from one answer. Treat it as one observation, repeat comparable prompts over time, and verify both accuracy and source quality.