Check your company’s visibility in AI
Free, with no credit card required.
A single answer is not a permanent position for your brand
On Monday, you ask about a project management tool and see your brand among the recommended solutions. On Tuesday, you enter a similar question and the list looks different. It is easy to interpret this as losing your position. First, however, check whether you are comparing the same test.
Perplexity combines information search with answer generation. Its Pro Search documentation describes model selection, searching multiple sources and synthesizing the information collected. The result therefore depends on the question, the available material and how the answer is prepared. This explains why a recommendation should not be treated as a reserved place on a list.
A brand missing from one answer is a reason to investigate. By itself, it proves neither a penalty nor a lasting decline in visibility.
Why recommendations can change
1. The question sets different conditions. “The best project management tool” and “a project management tool for a small agency costing up to PLN 100 a month” involve different criteria. Adding a price, market or use case can change which brands fit. A brand may be relevant to the first question but fail a condition in the second.
2. The conversation context changes. Perplexity’s documentation says follow-up questions can use earlier context. If you previously requested solutions for large companies, the next answer may still take that need into account. Compare fresh conversations and run contextual tests separately.
3. You use a different mode or model. Pro Search supports model selection and different source types. Changing a setting changes the test conditions. Record the mode, model and whether the answer comes from the user interface or an API.
4. The answer uses a different set of material. Perplexity describes searching and synthesizing information from the web. A new ranking, an updated review or an unavailable page can change the information behind an answer. This is one possible explanation for a particular omission; investigate it by comparing cited URLs.
5. The service itself changes. Model and feature updates can affect how answers are produced. Check provider announcements and your own settings records. Do not attribute every change to an update: matching dates are a clue, not proof of causation.
6. The wording of the answer changes. A generative answer is not a guaranteed, identical list of brands. Even similar material can be summarized differently. Without access to internal mechanisms, two screenshots cannot establish which cause determined the outcome.
A mention, a citation and a recommendation are different things
A brand may appear as an example, receive a negative assessment or be recommended for a particular task. Each situation has a different meaning. Likewise, a link to your domain may support a single fact without recommending a purchase.
Define your terms before measuring. This prevents you from declaring success simply because the system mentioned your company’s name.
| Signal | What you record | What you do not assume |
|---|---|---|
| Mention | The brand name appears in the answer | That the brand is recommended |
| Citation | The answer links to the brand’s domain | That the brand is the best choice |
| Recommendation | The brand is presented as meeting a need | That it will appear in every answer |
How to distinguish a fluctuation from a lasting decline
Start with a fixed set of questions reflecting customer situations. Include category questions, solution comparisons and purchase needs. Keep questions containing the brand name separate: they test awareness of the company, but do not show whether the system would suggest it independently.
Keep the question wording, language, market and settings identical. Start new conversations and record the full answer, date and sources. If you test personalization or conversation history, treat it as a separate scenario.
Repeat measurements on subsequent days. A weekly review can be a useful organizational starting point, but is not a universal statistical minimum. The number of observations needed depends on the number of questions, how often the brand appears and the size of the change you want to detect.
An illustrative example only: if the brand was recommended in 12 of 20 answers, the recommendation rate in that sample is 60%. In the next comparable set, 10 of 20 answers gives 50%. The difference is 10 percentage points. Such a small sample alone is insufficient to establish a lasting deterioration.
Compare the same question set and report the sample size. Adding many new, harder questions can lower the overall result without changing the answers to earlier questions. Analyze API and user-interface results separately; do not assume they reproduce the same context.
AI-generated conceptual illustration: regular measurement lets you compare a series of answers instead of judging a single result.
What to check when your brand is repeatedly omitted
First, compare the sources. Did your domain disappear, or just the brand name? Is the same website still cited, but a different article? Was a competitor recommended based on a new review? Changes in source URLs and changes in recommendations are related but distinct phenomena.
Next, check pages relevant to the question. Are they publicly accessible, do they contain the current offer, and do they clearly explain uses, limitations and price? Check server errors, redirects and access settings. A technical problem is a concrete reason to make a repair; you need not wait for a longer trend if a page is broken.
Check external material. Reviews, comparisons and industry publications can provide information about a brand. Do not assume, however, that one type of website always dominates Perplexity. Set priorities using the sources present in answers to your customers’ questions.
Finally, compare the scope of the decline. Does it affect one question, one category or many scenarios? Does it also appear in other systems? A recurring decline in comparable measurements warrants analysis. It still does not prove that a particular website change is responsible.
How to increase the chance of a relevant recommendation
Clearly explain who your offer is for and which uses it suits. Instead of a vague “complete solution for everyone,” provide specific features, conditions, limitations and examples of use. This helps both customers and the search system assess fit.
Update facts that influence choice: prices, availability, supported markets, integrations and purchase terms. Keep the brand name consistent across your own pages and external profiles. If you publish a comparison, explain the criteria and provide sources.
Develop reliable material answering questions where the brand performs poorly. This might be a use-case page, a compatibility explanation or accurate product documentation. Merely adding the word “Perplexity” to a text cannot replace useful information.
After making a change, observe a series of subsequent measurements. Record the dates of your actions to compare them with answers. An increase after publishing material is an observation; without an appropriate study, it does not prove that the material caused the increase.
How to use Semly to organize observations
A fixed question library and measurement history help establish where a brand appears regularly and where its presence is sporadic. Semly lets you analyze brand visibility, answers and sources, and compare results across periods and AI systems.
Treat a report as a starting point for selecting questions and areas needing attention. Separate changes in answer content from changes in sources, and always read a percentage result together with the measurement scope. Monitoring describes observed answers; it does not guarantee a permanent Perplexity recommendation.
Frequently asked questions
Does a missing brand mean a penalty?
You cannot establish this from a single omission. Check the test conditions, sources and whether the change recurs.
Does an identical question guarantee an identical answer?
Do not assume so. Besides question wording, settings, conversation context and the material used in the answer may matter.
Does citing a domain mean recommending the brand?
No. A citation may simply support a fact. Assess a recommendation by its content and its fit with the user’s needs.
How many measurements do you need?
There is no single number suitable for every brand. Collect a series of comparable observations, report the sample size and increase it if you want to evaluate small differences.
Can you guarantee a recommendation every day?
No. You can improve information availability and quality, and the fit between your offer and the questions. The service retains control over the content of its answers.
From a single answer to a decision
When your brand disappears from Perplexity, start by comparing the question, settings and sources. Only a series of consistent observations lets you assess whether the problem recurs. This helps you invest time in better information and real gaps instead of rebuilding your website after every different result.
Sources and scope of this explanation
Perplexity Help Center: What is Pro Search?
The documentation describes search, information synthesis, model selection and retention of conversation context. The possible reasons for omitting a particular brand and the diagnostic method presented here are practical inferences from that process, not a disclosed list of Perplexity ranking factors. The percentage example is illustrative; this article does not present results from a study of brands.
Share:
