Live, Cached or Disabled: What a Search Tool Setting Actually Controls

By Daoyu Guan · Focused review November 4, 2026

An AI answer can look like a judgment about the web when it is partly a consequence of how the application was configured. A live-search setting permits current web access through that tool; it does not guarantee that a particular page is discovered or used. Cached search uses saved content. A domain filter restricts that search path. Disabling built-in search does not establish that the entire application lacks current evidence: other tools, files or supplied context need their own review. GEO measurement therefore needs to record the retrieval environment, not just the prompt and final citation. Otherwise, teams may attribute a visibility change to content quality when the real cause is freshness, location, source policy, or the absence of search.

A documented search control, with a specific scope

OpenAI's Agents API web-search guide documents live, cached and disabled modes for built-in search, along with domain, location and context settings. Omitting the tool or selecting disabled turns off that built-in search path. Asking for a search in a prompt does not configure a missing tool. These documented controls are retained here; the correction concerns what can be inferred from them.

This is product documentation for one API, not a universal description of every AI answer engine. Its broader importance is conceptual. It shows that a conversational answer is produced inside a retrieval environment. Two applications can use the same model and receive the same question while operating over different evidence. Even the same application can produce different source sets after a configuration change.

For measurement, identify the application and the evidence paths it exposes, together with the model and time. An API's named control is not evidence that the consumer ChatGPT interface exposes or uses the same setting.

A more accurate model of an answer

A conventional visibility report often records four fields: platform, prompt, answer, and cited URLs. That is not enough for a causal diagnosis. A more complete representation is:

observed answer = f(model, conversation, tool availability, retrieval mode, corpus time, domain policy, location, context budget, answer policy)

This is an analytical checklist, not a measured causal equation or a claim that every product exposes these controls. Record known settings and mark the rest unknown.

This table does not imply that every consumer product exposes every variable. It establishes the variables an evaluator should control or mark as unknown.

Live search is not the same as immediate visibility

The word "live" can be misleading. It means the tool may access the current internet; it does not promise that every URL is already discovered, fetchable, parsed, selected, or cited. A newly edited page must still pass through several gates:

1. The system must know the URL or discover it from a query.
2. The page must be accessible to the relevant fetch path.
3. Its useful content must be present in a representation the system can process.
4. The page must be judged relevant enough to enter the working context.
5. Its claim must survive synthesis alongside other evidence.
6. The interface must decide whether and how to expose a citation.

A failed appearance at any one gate can produce the same visible result: no mention. That is why publishing at 10:00 and checking an AI answer at 10:05 is not a valid indexation test. It combines discovery latency, fetch behavior, retrieval selection, and generation into one observation.

Cached search creates a version problem

Cached mode describes the permitted source path; the label alone does not establish a latency, cost or reproducibility benefit. Hypothetical example: an invented software vendor changes a policy from 90-day to 30-day retention. Its current page shows 30 days while a saved copy contains 90. An answer based on the older copy can be faithful to that evidence and outdated for the buyer. These invented figures are not a Semvi policy or an observed customer incident.

The appropriate remedy is not to repeat the new number across dozens of pages. That can create internal contradictions and make future maintenance harder. Instead, maintain one canonical policy page, add an explicit effective date, update dependent pages, preserve a concise change record, and test whether the old value remains in feeds, PDFs, help-center mirrors, partner listings, or translated versions.

For volatile facts, a useful content block states five things together:

• the value;
• the entity or plan to which it applies;
• the effective date;
• the jurisdiction or audience;
• the canonical source.

That compact bundle is easier to retrieve and safer to quote than a promotional sentence detached from scope.

Domain allowlists change the meaning of "best source"

A domain allowlist constrains the search tool to which it is attached. It does not, on its own, define the entire application's evidence universe. A source excluded from that search path could still be present in an uploaded document, earlier conversation or another enabled integration. Claiming an application-wide approved-source boundary requires evidence about all relevant paths.

OpenAI documents computer use and MCP connections as separate integrations. Their existence is a reason to inspect the configured application, not evidence that a particular deployment has them enabled or that they can bypass its permissions. This article did not test any application's additional tools.

If the whole application is deliberately limited to approved material, a vendor explainer outside that corpus may indeed be unavailable. Establish that boundary from the actual configuration and observed tools rather than inferring it from one search filter.

This does not make the answer defective. Domain restrictions can be an appropriate safety and governance control. It does mean that a citation win inside one application cannot be generalized to all AI search.

GEO teams should classify tests into at least three source regimes:

Reporting all three as "AI ranking" erases the most important difference between them.

Location is part of the query even when the user never types it

The Agents API can receive country, region, city, and time zone as search-location hints. Other systems may infer geography from account settings, device state, interface language, or network information. A prompt such as "Which provider is available near me?" is incomplete without that context, but even apparently global questions can have local answers because product availability, regulation, language, and source access vary by market.

A defensible visibility study records both the explicit prompt and the environmental location. For international brands, it should use a market matrix rather than one global average. A source that performs well in London but is absent in Singapore is not unstable in the statistical sense; it may be responding to different source and product conditions.

Separate diagnostic probes from a discovery panel

When an answer appears outdated, register the diagnostic question and intended check before submitting it. The following steps describe possible investigations, not a required number of consumer retries. Use controls the product actually exposes; an unavailable control stays unknown.

Keep the original unbranded discovery question and answer intact. A prompt naming the company, prescribing a source or constraining a domain is a diagnostic probe. It does not earn natural-discovery or recommendation credit. A login wall, rate limit or unreadable answer belongs in the record; preserve the original attempt rather than repeatedly sending it until a favorable answer appears.

1. Run the baseline. Save the exact prompt, prior conversation, timestamp, account state, location, answer, links, and screenshots or raw output available through the product.
2. Ask a freshness probe. Use a fact that changed recently and has a clear official date. This does not prove live retrieval, but it can reveal an obviously stale path.
3. Label a source-directed probe. In a separate diagnostic record, ask the system to consult the official page. Compare the claim and displayed attribution. Keep this result out of the unbranded discovery denominator.
4. Constrain a documented search path. In an authorized API test, apply a supported domain filter and record other enabled evidence channels. This probes that configuration; it does not reconstruct a consumer application.
5. Change one location variable. Keep the prompt and all other settings fixed. Record whether the source set and conclusion change.
6. Use planned repeat dates. Retain the same question and comparable surface/context. No fixed cache lifetime is established here. A later readable answer is a later observation, not proof that a cache expired or a recommendation stabilized.

A bounded finding might be "the named-source probe displayed this page, while the unbranded answer did not cite it" or "these two configured API runs reported different policy dates." Those observations support a next investigation. They do not establish every hidden selection step or the cause of the difference.

Design content for both current and saved retrieval

Teams cannot control an application's cache policy, but they can reduce version ambiguity.

First, keep stable canonical URLs for durable facts. Replacing a policy page with a new path at every update fragments historical references. Second, place the last-reviewed or effective date near the claim it qualifies. Third, expose the decisive text in the page itself rather than only in an image, animation, or downloadable attachment. Fourth, maintain clear entity and product identifiers so an older variant is not confused with the current one. Fifth, make corrections explicit. In the hypothetical vendor example, a note stating the affected plan, effective date and changed retention period would make the policy transition inspectable. Verify those facts for the actual product before publishing a similar note.

Do not manipulate visible dates to simulate freshness. A changed date without a changed fact weakens auditability and can make readers distrust legitimate updates.

Measure the pipeline, not one output

A retrieval-aware GEO dashboard can use five stages:

Where a consumer interface hides tool traces, mark selection as unobserved rather than inferred. A citation proves that a source was exposed to the user. It does not reveal every source retrieved or every rule used to select it.

What this changes for GEO strategy

The immediate lesson is methodological. Content work, technical access, and measurement must share the same retrieval model. An editorial team can publish an excellent update while an evaluation team tests a cached corpus and reports failure. A developer can add a domain allowlist for safety while marketing interprets the resulting source concentration as a brand preference. A regional team can improve local pages while a global dashboard averages away the gain.

Before changing content, identify which part of the pipeline is failing. If a source cannot enter the candidate set, rewriting the introduction will not fix the problem. If the page is retrieved but the claim lacks scope or date, clearer evidence may help. If an answer is correct but uncited, report the correct claim and unresolved source use separately. Do not assume your page entered the context or that interface exposure is the only explanation.

Frequently asked questions

Does live search guarantee that a newly published page can appear?

Live access permits current web retrieval through the configured tool. Discovery, access, parsing, relevance and displayed citation remain separate observations.

Is a cached answer necessarily wrong?

A cached answer can agree with a stable fact. Check the source version when the decision depends on a changed policy, price, availability or event; the cache label does not establish which copy supplied a particular claim.

Can a prompt force an agent to search the web?

Not always. OpenAI's Agents API documentation says that web search must be configured as a tool; asking for search in the prompt does not enable a missing tool.

Should GEO reports compare models or applications?

Compare complete applications under recorded conditions. Model name alone does not identify retrieval mode, source restrictions, location, or interface policy.

What is the fastest test for a suspected stale answer?

Ask for the official source and effective date, then compare a live controlled-domain run with the original environment where possible. Treat the result as a diagnosis of that setup, not a universal platform finding.

Check the API contract before copying a setting

The Responses API web-search guide uses different controls from the Agents guide: it documents external_web_access for live versus cache-only access and states that the preview variant ignores that parameter. This is why a report must retain the API, tool variant and actual request. A matching label or provider name is insufficient to transfer a control between integrations.

Treat a published numeric limit with the same care. The current Agents explanatory guide and agent-update reference display different domain-list bounds. The review notes retain both observations; this article does not resolve them into a tested runtime quota. Obtain the contract for the exact endpoint and verify a bounded authorized request before promising a limit to a customer. No API request was run for this article.

For a monitoring purchase, agree whether the deliverable samples a consumer web interface, an API configuration or a private corpus. The Semvi by Xindar scope reference separates product descriptions and confirmations from account behavior still requiring acceptance. Xindar's first-party research record keeps ordinary consumer observations distinct from source-directed probes. Neither page proves a hidden retrieval mode or constitutes an independent customer result.

Sources and evidence boundary

This focused review used the current official Agents web-search guide, Responses web-search guide, agent-update reference, computer-use documentation and MCP connection documentation linked at the relevant claims. They establish those documented controls and integration boundaries, not the private configuration of ChatGPT, Semvi or another application. The framework, diagnostic procedure and content-maintenance advice are editorial recommendations. The retention-policy example is invented. No model API, consumer question, account export or runtime-limit test was executed for this revision.