The OpenAI Roadmap in 2026: What Shipped and What Slipped

This page used to summarise a leaked OpenAI strategy document. That framing has not aged well, because most of what was leaked has now either shipped, changed shape or quietly disappeared. So it has been rebuilt around something more useful: a dated record of what OpenAI said it would do, set against what it actually did.

Last reviewed 3 August 2026. Every date below is sourced. Where a figure comes from secondary reporting rather than a primary document, it says so.

The short version

In the last eighteen months OpenAI shipped its advertising business early, fast and across seven markets. In the same period it pulled back from owning the checkout, released no GPT-6, and moved its first consumer device to 2027.

That split is the whole story. The parts of the roadmap that monetise attention arrived ahead of schedule. The parts that required OpenAI to own the transaction or the hardware slipped. If you are deciding where to spend effort as a brand, that pattern tells you which surfaces are worth competing for.

What OpenAI promised, and what actually shipped

This is the section every other roadmap article is missing. Anyone can restate an announcement. A track record of promise against delivery is harder to argue with, and it is the only honest basis for predicting what happens next.

Sources for the rows above: Forbes, 16 January 2026 on the ad announcement; Criteo's own press releases of 2 March and 22 June 2026 on the ad-tech partnership; CNBC, 24 March 2026 and Forrester on the checkout retreat; Modern Retail on the Shopify merchant count; the GPT-5.5 release record; and MacRumors, 10 February 2026 on the device delay. The Shopify merchant numbers are secondary reporting of executive remarks, so treat them as directional.

Why ads were the real pivot

The advertising launch gets covered as a reversal of a personal position. The more useful way to read it is as arithmetic.

Roughly 5% of ChatGPT's weekly users pay for a subscription. The remaining 95% generate the engagement, the compute cost and the reach that makes the platform interesting to advertisers, while contributing nothing to revenue directly. Any business with that shape and OpenAI's infrastructure commitments ends up selling attention. The Harvard remark was made in May 2024, before the scale of that commitment was public.

We have early access to OpenAI's ads platform and buy inventory on it, which is where most of the detail below comes from.

Three mechanics matter for anyone planning around this:

• Ads render outside the answer, below it, and OpenAI states plainly that they do not influence the answer you get. The organic answer and the paid slot underneath it are two separate competitions.

• Placement is decided by conversational relevance, and the landing page, ad copy and context signals are all scored together. Incoherence between them costs you money in the auction.

• Serving is limited to the free and Go tiers. Your Plus and Pro customers see the answer with no ad beneath it, so for those users the only surface that exists is the organic citation.

That last point is the commercially important one. An ad-funded answer layer does not make organic AI visibility less valuable. It splits your audience into a segment you can pay to reach and a segment you can only earn.

Where OpenAI did deliver: retrieval

Under the noise about models and hardware, the retrieval side of ChatGPT shipped and matured. It is now the part of the product most likely to decide whether a customer ever hears your brand name.

• ChatGPT cites roughly 15 sources per response where Gemini cites about 3 (Semrush AI Visibility Index, 126 million US prompts, January to April 2026). More citation slots means more room for a challenger brand to be one of them.

• Its citation mix is its own. Wikipedia at 13.15% and Reddit at 11.97% lead, with OpenAI's own properties at 6.21% (Similarweb, April 2026, around 600,000 citation events). That profile looks nothing like Google's AI Overviews, where YouTube and Reddit dominate.

• OAI-SearchBot is the crawler that governs whether you can appear in ChatGPT search at all. Blocking it removes you from the surface completely. OpenAI documents this.

• OpenAI's crawl-to-referral ratio was about 1,700 to 1 in June 2025 (Cloudflare). The traffic economics of being read by ChatGPT are unlike being crawled by Google, which is precisely why the citation itself has to carry the value.

What this changes for your brand

Three things follow from the record above, and they are the same three we work through with clients.

1. Compete for the answer, because it is the surface that keeps proving durable

OpenAI has now retreated from owning the checkout once and moved its hardware by a year. The answer itself has only become more central across the same period. When a company keeps pulling back from the edges of its product and doubling down on the middle, the middle is where you invest.

This is measurable work with a normal timeline. A Portuguese childcare marketplace went from zero ChatGPT visibility to first place on its target prompts in forty days, with its first AI-sourced paying customer on day 38 (Baby Sisters case study). A US recovery centre in a heavily regulated market moved its high-intent prompt set from 0% to 67% visibility, with AI visitors converting into enquiries at five to six times the rate of Google organic traffic (case study).

2. Budget for two audiences, because ads only reach one of them

Ads serve on the free and Go tiers. Everything above that sees the answer alone. Any plan that treats ChatGPT ads as full coverage of ChatGPT is undercounting, and the segment it misses skews towards the users paying for software.

3. Stop spending on the two levers the evidence has now closed

Both of these appeared in earlier versions of this article. Both are now contradicted by the primary evidence.

• Schema markup as a route to AI citation. Google's own generative-AI optimisation guide states that structured data is not required and that there is no special markup to add (Google, updated 10 July 2026). Ahrefs tested 1,885 pages that added JSON-LD against roughly 4,000 matched controls in May 2026 and found a null result. Schema remains worth having for indexing and rich results. It is not an AI-citation lever.

• llms.txt. Ahrefs examined server logs across 137,210 domains in June 2026 and found that 97% of published llms.txt files received zero requests in May 2026. Of the 3% with any traffic, AI retrieval bots accounted for 1.1%.

What to watch between now and the end of 2026

• Whether a genuine GPT-6 arrives this year. As of August 2026 there is no architecture paper, no parameter count, no pricing and no date.

• Whether ad serving extends above the free and Go tiers. That single change would redraw the split described above.

• Whether OpenAI returns to owning the transaction, or leaves commerce as discovery inside ChatGPT and purchase on the merchant's site.

• Cloudflare's new AI bot defaults, which take effect on 15 September 2026 for new domains and can silently block search crawlers for anyone who ever ticked the legacy "block AI bots" option.

How we track this

We monitor prompt-level visibility across ChatGPT, Gemini and other models for our clients, measured through APIs. That is an honest instrument with a known limit: an API query is not identical to a real logged-in ChatGPT session, so we report presence and share of voice over time and make no claims about ranking positions inside a consumer chat. If you want to see where you currently stand, the free visibility check runs against your own market.