I read all 840 of OpenAI’s open jobs. Job ads often reveal more about what a company is building next than its press releases do.
Companies usually talk publicly about what they have already built.
Job listings are different.
Before a role goes live, someone has normally approved the budget, the team and, at least broadly, the work that person will be doing.
That makes a careers page surprisingly useful if you want to understand where a company is heading.
So I went through OpenAI’s careers page.
Not a sample. All 840 open roles listed on 29 September 2026, with a closer look at the 116 jobs connected to how ChatGPT finds, retrieves and chooses information. Three-quarters of those 840 roles had been posted since June.
And the picture that emerges is much bigger than another model release.
OpenAI appears to be building its own search infrastructure, expanding advertising, pushing deeper into science and government, preparing consumer hardware, growing an agent and plugin ecosystem, and building the financial and operational machinery of a much larger company.
More importantly for anyone working in SEO, GEO or AI visibility, the listings reveal more about how ChatGPT search itself is evolving.
Here is what I found.
First, can job postings actually predict what OpenAI will do?
There is a reasonable objection here: job descriptions are partly recruitment marketing.
Teams describe ambitious plans. Some projects get delayed. Some disappear altogether.
So before treating the current board as a roadmap, I looked at whether older postings lined up with things OpenAI later announced.
They frequently did.
Roles appeared around initiatives such as ChatGPT Work, hardware, security products, life-sciences tooling, consulting, data centres and robotics before those projects became widely visible externally.
The careers board often appears to run weeks or months ahead of the public announcement cycle.
That does not mean every listing predicts a launch.
It does mean the job board is worth watching.
And it changes quickly.
Three weeks before this analysis, the word “ranking” appeared only three times across the board, all inside advertising roles. By 29 September, it appeared six times, including twice on the search side.
One posting from 25 September materially changed what the search architecture appears to look like.
Part One: How ChatGPT May Find and Choose Its Answers
1. OpenAI is building its own search infrastructure
One of the clearest signals is compensation.
A search researcher role listed a salary range of $380,000 to $850,000, while another search position went as high as $585,000.
The work described includes building systems that transform and index enormous corpora of information.
In practical terms, that is search-engine infrastructure: collecting large amounts of information, organising it and making it retrievable quickly.
The surrounding roles reinforce that interpretation.
There is a crawling team responsible for web crawling and GPTBot services, infrastructure powering vector search, and a newer role that explicitly includes freshness as something the search system will be measured against.
For website owners, that matters.
ChatGPT search is increasingly something that should be understood as its own retrieval environment rather than simply a layer sitting on top of somebody else’s results.
2. The ChatGPT search pipeline now has named stages
This may be the most interesting search-related listing in the entire dataset.
A 25 September role for Search Product Infrastructure describes services coordinating:
search classification, retrieval, ranking and model inference.
That gives us a much clearer mental model of the process.
A query can be:
Classified → retrieved → ranked → passed to the model for inference
The ranking stage is especially important because earlier evidence from the careers board made ranking appear much less prominent.
But the first stage may be just as important.
Search classification happens before retrieval.
That suggests the system first determines what kind of request it is dealing with, which could influence the data source, vertical, tool or search behaviour used afterwards.
The same posting also says OpenAI is building source attribution and access controls into the search infrastructure itself.
For GEO and AI-search measurement, classification may become one of the most important parts of the system to understand.
3. A lot of search behaviour appears to be learned
Across the 840 listings, the language around search is striking.
The file records:
- Retrieval: 56 job listings
- Evals: 45
- Memory: 42
- Graders: 29
- Indexing: 11
- Ranking: 6
There is clearly a ranking stage.
But the organisation described in the job board does not look like the traditional search-engine model of large relevance teams, quality-rater programmes and public ranking guidelines.
Instead, OpenAI repeatedly talks about data, graders, training environments, evaluations and feedback loops.
One product-management posting describes a closed loop that turns product usage and explicit feedback into datasets, experiments, training priorities and launch decisions.
That distinction matters.
Some of ChatGPT’s search behaviour may change because the model itself changes, rather than because somebody manually adjusts a conventional ranking factor.
I have already seen evidence of this in my own testing.
In July, ChatGPT ran multiple rounds of searches and opened pages. A month later, the search behaviour had become narrower: fewer rounds, no page opens in the August run, and some results were represented by 200-character rendered-page extracts.
Same underlying task. Noticeably different retrieval behaviour.
That is why testing AI visibility only once is increasingly risky.
4. Search is splitting into verticals
Another search listing talks about infrastructure for individual “Search product verticals.”
That language matters.
It suggests ChatGPT search is not necessarily becoming one single universal search product. Different categories can develop their own retrieval behaviour, tooling and datasets.
Shopping is already a useful example.
Instead of relying entirely on ordinary crawling, merchants can provide structured product feeds. Health operates in a different environment again, while finance is becoming part of its own product experience.
That means a single company-wide “AI visibility score” will become less meaningful.
Visibility for a shopping query and visibility for a health or B2B software query may eventually be measuring different systems.
5. Personalisation makes rankings less universal
OpenAI is also hiring heavily around memory, personalisation and proactive assistants.
The stated direction is to evolve ChatGPT from a chatbot into a highly personalised assistant, with memory carrying context across ChatGPT and other agentic products.
That creates an obvious measurement problem.
If two users ask the same question but the system knows different things about each of them, there may no longer be one definitive answer—or one definitive “position”—to track.
AI visibility measurement will increasingly need controlled testing environments, such as memory-off sessions or fixed profiles, if results are meant to be comparable.
Part Two: The Businesses OpenAI Appears to Be Building
6. Advertising inside ChatGPT is becoming a serious business
Roughly 45 of the analysed roles relate to advertising.
The listings cover ad buying, ad selection, sales, integrity and multiple creative formats including text, image, video, native, conversational and interactive ads.
There is another detail website owners should pay attention to.
OpenAI now distinguishes between multiple crawlers, including OAI-SearchBot, GPTBot, OAI-AdsBot and ChatGPT-User.
OAI-AdsBot is particularly interesting because the source material describes it as reading advertiser landing pages both for safety and to determine relevance.
That means landing-page content may influence more than organic AI visibility.
It can also become part of the paid advertising system.
7. Life sciences is becoming a full product category
One of the newer signals is the emergence of Life Sciences as a dedicated department.
Roles include engineering and product leadership for Rosalind Workbench, along with delivery roles serving pharmaceutical companies, clinical research organisations, drug discovery and related workflows.
The Workbench is described as bringing scientific tools, data sources, biological file viewers and life-sciences workflows together.
The longer-term goal described in the postings is even more interesting: teams of agents working together across these domains.
This looks less like “another AI model for science” and more like specialised vertical software.
If the approach works, it is easy to imagine the same organisational pattern appearing in other industries.
8. OpenAI’s consumer device is moving toward certification
The Consumer Devices organisation had 27 jobs in the analysed board.
Two listings stand out.
One is for a wireless regulatory engineer working on SAR compliance from prototypes through testing, certification and production.
Another focuses on on-device inference, including quantisation, runtime integration, memory constraints, compute and energy budgets.
That combination suggests work beyond an early laboratory prototype.
It also gives us a clue about the product itself: at least some AI processing is intended to happen locally on the device.
That could improve latency, enable some offline behaviour and create a stronger privacy story.
9. OpenAI is moving into manufacturing and outcome-based commercial models
A newer Product Design & Manufacturing vertical is aimed at customers working from product engineering through factory operations.
The stated use cases include shorter development cycles, improved factory uptime and higher output or yield.
Target sectors include automotive, industrial equipment, agriculture and consumer electronics.
The commercial structure is also notable.
Other roles mention terms such as revenue sharing, royalties and outcome-based pricing.
That pushes OpenAI closer to consulting firms, industrial software vendors and implementation partners—not just AI-model providers.
10. Government is becoming a dedicated customer segment
The source identifies 16 roles inside OpenAI for Government.
They include programmes for sensitive public-sector use cases involving cyber and biology, along with a delivery role supporting the U.S. Department of Energy, national laboratories, NASA and the National Science Foundation.
The notable organisational signal is not simply that governments use OpenAI products.
It is that government customers are getting dedicated programmes, delivery roles and access structures for specialised use cases.
11. Agents may eventually bypass conventional search traffic
Another cluster of jobs is focused on assistants that can operate browsers, desktops and web interfaces directly.
One listing uses the phrase “hybrid agentic/interactive web browser,” while other roles focus on training models to operate computer interfaces and building always-on virtual coworkers.
Alongside this sits a broader plugin and connector ecosystem, including work around MCP, permissions, publishing and discovery.
This matters enormously for websites.
In traditional search:
Crawler → search results → human → website
With agents, the path can become:
Agent → website → action
The visitor may no longer be a human.
And optimising for that environment will require more than conventional SEO.
Forms need to work. Important information cannot be hidden behind unnecessary interaction. Login requirements, pop-ups and frontend behaviour become machine-accessibility questions.
12. Enterprise support is becoming its own paid layer
Five Dedicated Support Engineer roles appeared across San Francisco, Dublin, Tokyo, Singapore and Toronto.
The roles support premium customers with account-specific documentation, proactive monitoring and incident management.
That is another sign of OpenAI moving from “AI product company” toward enterprise infrastructure provider.
Part Three: The Company OpenAI Is Becoming
13. The finance organisation is being built for a much larger company
OpenAI is hiring for M&A integration, financial reporting, technical revenue accounting, internal controls and work associated with regulatory and board reporting.
The signal is not that a specific corporate transaction is guaranteed.
It is that OpenAI is building the finance infrastructure required to handle acquisitions, more complex reporting and a substantially larger organisation.
14. OpenAI wants a role in defining AI assurance standards
Two roles are especially revealing.
One covers Frontier Assurance, including third-party assessments, embedded auditing and a misalignment disclosure process.
Another focuses on standards involving mental health, youth safety, age assurance, reliability, efficacy and privacy across organisations such as ISO, NIST-related groups and other industry bodies.
The takeaway is broader than compliance.
OpenAI is not only preparing to operate under emerging AI standards. It is also investing in people who will participate in how those standards are defined.
15. Compute remains fundamental to everything—including search
Scaling is one of the largest organisations represented on the board.
The jobs span electricity, gas, utilities, land, community engagement, custom chips, inference infrastructure and networking.
This has a direct relationship with search.
Searching more sources, opening more pages and running more inference costs compute.
That means the depth of an AI-search experience is partly an economic decision.
My July-to-August experiment showed exactly the kind of behavioural difference that makes this important: fewer search rounds, fewer domains and no page opens in the later run.
For AI visibility practitioners, infrastructure economics may matter more than we traditionally assume in SEO.
16. OpenAI is building country-level commercial teams
Recent listings include engineering, strategy, partnerships, marketing and commercial roles across the UAE, MENA, the UK, France, Germany, Australia and New Zealand, Latin America and EMEA.
This looks increasingly like country-by-country commercial expansion rather than a company operating everything centrally from San Francisco.
What OpenAI is not hiring for is just as interesting
Sometimes the gaps tell you more than the openings.
There is no obvious equivalent of a Search Console for publishers.
There is no substantial publisher-relations organisation.
There is no public-facing system explaining why a website was or was not cited.
And while ranking now appears explicitly inside the search pipeline, the job board still does not show a traditional search-ranking organisation filled with relevance engineers, raters and public quality guidelines.
There is, however, one slightly ironic exception.
An OpenAI marketing role explicitly mentions growing organic discovery through SEO, content and GEO.
So OpenAI is hiring for generative engine optimisation internally while the rest of us are still trying to work out exactly how its own generative engine chooses sources.
What I think happens next
The job board does not let us predict exact launch dates.
But it does show where money and headcount are moving.
Over the next several months, I would watch advertising formats, search verticals, industry-specific products and changes in ChatGPT’s retrieval behaviour particularly closely.
Over a longer window, personalisation will make universal AI rankings harder to measure, agents will create a new category of machine-driven website traffic, and structured data feeds may become increasingly important in commercial verticals.
Consumer hardware, specialised scientific products and infrastructure services appear to be moving on separate but parallel tracks.
The larger pattern is simple:
ChatGPT is becoming less like one product and more like a collection of search, agent, commerce, enterprise and vertical-software systems sitting on top of the same AI platform.
That changes how we should think about visibility.