Grok’s X Search can retrieve current posts, profiles and threads while a response is being produced. That access is unusually valuable during a breaking event because it reaches first-party statements, eyewitness material and public reaction inside one source. It also creates a source-selection problem: the search tool can filter handles and dates, but it does not decide which accounts deserve evidential authority.[1]
xAI exposes allowed and excluded handle lists, date ranges and optional image or video understanding. These controls define the searchable population. They cannot establish that an allowed account is truthful, that a popular post is original or that a quoted post preserves the context of the source it repeats.
X Search controls and what they actually control
| Control | What it changes | What it cannot prove |
|---|---|---|
| Allowed handles | Restricts results to as many as 20 specified accounts | That every allowed account is authoritative or independent.Source 1 |
| Excluded handles | Removes as many as 20 accounts | That remaining posts are representative.Source 1 |
| Date range | Limits posts to inclusive start and end dates | That a post describes an event that occurred in that period.Source 1 |
| Image understanding | Lets the agent analyse encountered images | That the image is authentic, current or correctly captioned.Source 1 |
| Video understanding | Lets the agent analyse encountered X videos | That the video is complete, original or geolocated.Source 1 |
Filtering by account is an editorial policy
A useful search policy distinguishes account roles before the model starts. A company account can establish what the company announced. A regulator can establish what it ordered or published. A named reporter may contribute verified reporting. An eyewitness can contribute a direct observation. An anonymous aggregator can help discover a lead but rarely settles the underlying fact by itself.
Those categories are not a ranking of honesty. They define what each source is in a position to know. A company is authoritative about its own announcement but not necessarily about the success of its product. An eyewitness may know what occurred in one place but not the cause. The agent should preserve the account, timestamp, post relationship and original claim rather than turning all retrieved text into one undifferentiated consensus.
Real-time search now has an item-based bill
xAI says that from 21 September 2026, X Search is billed at US$5 per 1,000 posts fetched and US$10 per 1,000 profiles fetched, in addition to model-token charges. Parent and quoted posts count when returned. Cost therefore depends on the retrieval path the agent chooses, not only the user’s final prompt.[2]
For an arithmetic illustration, 200 fetched posts cost US$1.00 and 20 fetched profiles cost US$0.20 under those rates, before reasoning, input, completion, image or video tokens. The example is not a forecast of a normal query. It shows why a short answer can sit on top of a much larger and variable evidence-gathering operation.[2]
A source record worth preserving
| Measure | Value | Context |
|---|---|---|
| Account identity | Handle plus stable profile identifier | Handles can change; the identity used for the claim should remain reconstructable.Source 1 |
| Post identity | Post URL or ID | Required to revisit the exact public statement.Source 3 |
| Time | Post timestamp and search-access time | A later edit, deletion or reply can alter interpretation.Source 1 |
| Relationship | Original, reply, quote or parent | Quoted text may depend on a thread or source post.Source 2Source 1 |
| Source role | First party, reporter, eyewitness, expert or aggregator | An editorial classification that helps decide what the post can establish. |
| Corroboration | Independent source or primary document | A separate evidence chain is stronger than several posts repeating one claim. |
Virality and authority are different signals
A model can see that a claim is repeated frequently. Repetition can represent confirmation, coordinated amplification, a widely copied error or thousands of reactions to one original post. Counting mentions without tracing the source tree risks converting popularity into evidence.
For high-consequence reporting, X Search should produce a lead set and citation trail, not the final standard of proof. Ask the agent to separate first-party claims from independent confirmation, find the earliest identifiable source and state what remains unresolved. Then compare the result with documents, official notices or reporting outside X.
Grok’s access to X is a major strength when the information is genuinely happening there. The limitation is not speed. It is that source authority is a policy decision the search tool cannot make on the user’s behalf. A trustworthy implementation makes that policy explicit and preserves enough evidence to challenge the model’s synthesis.
Sources & further reading
- X Search xAI Developer Docs · Accessed 10 Oct 2026
- API pricing xAI Developer Docs · Accessed 10 Oct 2026
- Citations xAI Developer Docs · Accessed 10 Oct 2026