SearchGPT: a citation is not a visit to your website

Being named in an answer and receiving a reader are different commercial outcomes. A publisher can value both, but should not use evidence of one to claim the other. That is the most useful business lesson from SearchGPT's July 2024 announcement: source attribution deserves its own measure, while visits and the actions that follow need separate evidence.

This is a retrospective written in November 2026. It does not estimate how much traffic the prototype sent, or how much it displaced. The argument concerns the decisions a publisher can make when an answer service promises visibility and the publisher still has to fund the work being cited.

What the prototype promised

On 25 July 2024, OpenAI announced SearchGPT as a temporary prototype for a small group of users and publishers. It described named attribution, inline links and a sidebar of sources. It also said search inclusion was separate from training: sites could appear in results while opting out of training its foundation models.

These are attributed statements about the product's design and announced controls. They do not establish referral performance. A visible source link creates an opportunity to visit. It cannot establish that a reader followed it, returned later or paid for anything.

The later product has its own chronology. ChatGPT search launched on 31 October 2024, initially for Plus and Team users and SearchGPT waitlist users. That page also carries later access updates dated 16 December 2024 and 5 February 2025. Broader access belongs to those later stages, not the July prototype. Reading the current page without separating its dates would exaggerate the audience available in July.

Measure the outcome your business needs

An illustrative specialist publication might sell subscriptions to original analysis. A citation can associate its name with expertise, yet leave the subscription relationship untouched. An illustrative service business might publish instructions mainly to help prospective customers assess its competence. A small number of relevant visits could matter more to it than widespread mention. These examples describe different business models, not observed customers or measured results.

Start by naming the desired action. For the publication, that may be a subscription. For the service business, it may be a qualified enquiry. A count of citations is an intermediate observation whose value depends on that objective. Reporting it as a substitute for the objective hides the question the owner actually needs answered.

Keep an evidence chain rather than a single visibility score. Record where and when a citation was observed, which page it referenced and what question produced it. Separately examine attributable visits where referral information is available. Then examine the relevant on-site action using whatever appropriate measurement the site actually has. Missing links in that chain should remain unknown.

This proposed method has limits. An answer sample is not a census of what every user sees. A visit whose referral information is missing cannot confidently be assigned to one service. A later enquiry may have several influences. Precision about those limits makes the report more useful than a neat total assembled from incompatible observations.

A source can help without winning the click

The strongest counterargument is that referral traffic is too narrow a definition of value. A reader might recognize the publisher's name in a cited answer and return directly later. For a brand seeking recognition, attribution may have value even without an immediate visit. It would be a mistake to declare that value zero merely because the click is absent.

Accept that possibility without converting it into an established outcome. Record recognition as a hypothesis and choose evidence suited to it, such as a relevant audience study if one is feasible. Do not quietly add imagined later visits to the referral column. The difference between an unmeasured benefit and no benefit matters, especially when choosing whether to fund more content.

For a paid publication, the trade can also be unfavourable. If the answer satisfies the reader's immediate need, the existence of a citation may leave little reason to open the source. That is a plausible mechanism, not proof that SearchGPT reduced anyone's revenue. Its significance depends on the publication's product and on observed reader behaviour.

Keep access and investment decisions separate

The July announcement's distinction between search and training helps publishers avoid bundling two choices. Allowing discovery does not, in that announcement, require opting into foundation-model training. But a documented control is only one input to a commercial policy. It does not promise selection, accurate attribution or a sustainable referral relationship.

This article's decision comes after access: what evidence would justify spending more on material intended to be found through answers? It is different from a crawler-blocking policy, which determines which requests a site wants to admit. Both decisions matter, but they should produce different records and different success criteria.

A practical starting point is a small, explicit observation period using existing content. Avoid rewriting a whole archive around a few sampled answers. If an attributed visit arrives, examine whether the destination helps the reader continue: clear authorship, useful supporting material and an appropriate next step. None of those changes guarantees more citations or a ranking gain.

Website upkeep can include this work when its scope is agreed through software maintenance. The useful deliverable is a bounded account of observed mentions, attributable visits and business actions, with the gaps intact. SearchGPT made the relationship between an answer and its sources more visible. Publishers still need evidence about the relationship between that visibility and the work they can afford to produce.