AI answers with and without web search: how they affect your score
Why it matters whether an AI assistant searched the web before answering, how often assistants search, and how that changes your visibility score.
An AI assistant can answer a question in two ways: from what its model learned during training (its own knowledge), or after searching the web and reading pages (an answer with web search, also called a web-grounded answer). The two can name different brands, only the second cites live sources, and your visibility score can move when the mix between them changes, even if nothing about your brand did. This article explains the difference and how Geovium handles it.
What is the difference between the two kinds of answers?
| From the assistant's own knowledge | With web search | |
|---|---|---|
| Where the information comes from | Training data, up to the model's cutoff | Pages retrieved at answer time, plus training |
| How current it is | As of the training cutoff | Can reflect recent changes |
| Sources | None retrieved; addresses it writes may be outdated or invented | Citations returned by the provider |
| What changes it | New model versions | Your pages, third-party pages, search indexes |
| Cost to measure | Model usage only | Model usage plus search fees |
Neither kind is "the real answer". People get both, depending on the assistant, the question and the settings.
Do AI assistants always search the web?
No. Perplexity is built around search and searches for every answer. For ChatGPT, Gemini, Claude and Grok, web search is a tool. In the consumer apps the assistant usually decides whether to use it, and users can ask it to. Through the APIs, the developer can offer the tool and, in some APIs, require it.
In our own test scans of ten questions each:
- Gemini searched the web in 6 of 10 answers, with search enabled. In earlier checks, before Geovium explicitly asked Gemini to search, it rarely searched on evergreen questions, meaning questions whose answer does not obviously depend on recent events.
- ChatGPT, for which Geovium requires the search tool, searched in 10 of 10 answers.
These are small tests, not a representative study, but they show the point: with the same settings, one assistant can search every time while another decides question by question. Claude and Grok also decide for themselves when the search tool is available.
How does web search change what the assistant says about you?
- Which brands appear. A searched answer draws on pages that exist today, so newer brands, recent reviews and new comparison articles can appear. An answer from the model's own knowledge favors brands that were well described in its training data.
- Which facts are used. Prices, features and availability are more likely to be current when the assistant searched.
- Whether sources are cited. Only searched answers come with citations, so only they can send readers to your website.
- How local the answer is. For questions tied to a city or a country, search can bring in local pages and listings that the model's own knowledge may not cover in detail.
How does it affect your Geovium visibility score?
Every complete answer is scored with the same formula, whether or not the assistant searched: mention (40 points), mention in the first paragraph (15), your website named or linked (15), list position (up to 30), positive tone (10) and trust (10), capped at 100. The difference lies in the website points.
- Website points need evidence. They count when your site appears in the citations the provider returns, or in the answer's source list when the answer was produced with web search.
- Without a search, a link is only a mention. If the model writes your web address from memory, or names your domain in plain text, that counts as a mention of your brand but earns no website points. Addresses written from memory are often wrong or invented, so they are not treated as evidence.
- Other signals are unaffected. Mention, early mention, position, tone and trust can be earned in either kind of answer.
Two consequences follow. First, if an assistant searches less often in one period, your link rate and your website points for that assistant can drop even though nothing changed in how the web describes you. Second, the cited-sources report only has data from answers where a search ran; if no answer searched in a period, Geovium says so instead of showing zero.
Geovium records with every answer whether web search was on and whether the provider reported that a search actually ran. The methodology page describes the full scoring.
How can you tell whether an answer used web search?
In the consumer apps, the clearest sign is a list of sources or links attached to the answer. An answer without any links was probably written from the model's own knowledge, but not always, and the reverse is not reliable either: a model can write links from memory that look like sources.
For measurement, rely on what the provider reports rather than on the answer text. The APIs report, for each answer, how many web searches ran, for example as a count of search queries or of search tool calls. That count is what Geovium records. When a provider reports no count for an answer produced with web search on, the answer is treated as searched; this is always the case for Perplexity, which searches for every answer. An answer where web search was on but the provider reports zero searches is treated as an answer from the assistant's own knowledge.
How should you read a score change?
Before crediting or blaming your own work, check these in order:
- Did the share of answers with web search change for that assistant?
- Did coverage change? Fewer complete answers means a smaller, noisier sample.
- Was there a model update? Providers change models without much notice, and answer patterns can shift overnight.
- Are you comparing the same questions, with the same filters and the same period length?
- Only then look at your content changes and at what competitors did.
An illustration, not data: say Gemini searched in 6 of 10 answers last week and in 3 of 10 this week. Even if it mentions you just as often, fewer answers can earn the website points, so your Gemini score can fall. The fix in that case is not a new landing page; it is reading the change correctly.
Should you measure with or without web search?
Both are legitimate; they answer different questions.
- With web search, Geovium's default, is closer to what people see when the consumer apps search the web, reacts sooner to your content work, and shows the sources behind each answer.
- From the assistant's own knowledge shows what the model has absorbed about your brand. It changes slowly, mostly with new model versions, which makes it a measure of long-term brand strength rather than of last month's work. It also costs less to measure.
In Geovium, workspace admins can choose the mode for each assistant in the AI assistant settings; Perplexity always searches. Every result remembers which mode produced it, changes apply to future scans only, and results from the two modes are not directly comparable. Pick a mode for each assistant and keep it, so your trends stay meaningful.
What should you do differently for each kind of answer?
For answers with web search, the levers are the ones that affect what search finds today:
- crawlable pages with the text in the HTML, and search crawlers allowed;
- pages that answer each important question in the first paragraph;
- a presence on the third-party pages that are already cited for your questions;
- current facts and visible update dates.
For answers from the model's own knowledge, the lever is slower: a broad, consistent and accurate presence across the web, which is the material future model versions learn from. Clear descriptions of what you do, independent coverage and consistent facts help with both kinds.
Our guides on which sources AI assistants cite and why ChatGPT may not mention your brand go deeper into each lever.
Frequently asked questions
What is a web-grounded AI answer?
It is an answer the assistant wrote after searching the web and reading pages, rather than from its training knowledge alone. Grounded answers can reflect recent changes and come with citations to the pages used.
Why didn't Gemini search the web for some of my questions?
Gemini decides for itself whether a question needs a search, even when search is enabled. In our test it searched in about 6 of 10 answers. Evergreen questions are the most likely to be answered from its own knowledge.
Do answers without web search count in my score?
Yes. They are scored with the same formula, but they cannot earn the 15 website points, because a web address written from the model's memory is not evidence that your site was used.
Are results with and without web search comparable?
Not directly. The two modes draw on different information, so Geovium records which mode produced each result. Compare trends within the same mode.
Which AI assistants always search the web?
Perplexity searches for every answer. ChatGPT, Gemini, Claude and Grok treat search as a tool; whether they use it depends on the settings and often on the assistant's own decision.