Company X-Ray report
OpenAI
openai.com · Artificial intelligence · San Francisco, CA
You own the category’s attention but not its definition – the explaining is being done on other people’s pages.
- Generated
- 12 March 2026
- Runtime
- 6m 24s
- Sources read
- 1,978
- Sections
- 17
Company Intelligence Score
This is a demonstration of the LEMANS LABS report format applied to a well-known public subject. It is not a real assessment of OpenAI, contains synthetic measurements produced for illustration, and is not affiliated with or endorsed by OpenAI. Your own report analyses your own public record.
Executive summary
Everyone can name you; far fewer can finish the sentence. Across the 1,978 sources we read, the company name appears in almost every general conversation about the category, but the description attached to it slides between “the ChatGPT company”, “an AI research lab”, and “the API everyone builds on” depending on who is speaking. Recall is solved. Definition is not, and definition is what a buyer repeats to a colleague.
The second structural fact is the ratio between earned and owned. Our crawl found 1,040 pages across your own domains against roughly 35,000 indexed third-party pages that explain what you do. When someone searches “what is a context window” or “fine-tuning vs prompting”, the page they read is written by a tutorial site, a newsroom, or a competitor’s documentation team. The vocabulary of your own category is being set by writers you do not employ.
The third gap is the material a procurement reviewer needs. We counted 23 pages covering security, data handling, compliance posture and model behaviour against 41 product and research pages – a developer arrives to a full library and a legal reviewer arrives to a summary. None of this is an awareness problem. The work is definition, owned explainers on the words that lead to purchase, and a trust estate sized to the audience already showing up.
Company overview
OpenAI is the most-named company in its category and the least consistently described. Founded in 2015 as a research lab, it became a household name through one consumer product and now runs a developer platform, a business tier and a research programme under the same word. The gap between recall and description is the subject of this report.
The product name outruns the company name
In general-audience sources, ChatGPT is referenced without the company attached in a clear majority of mentions. That is a strong consumer outcome and a weak corporate one: the brand equity accrues to a product line rather than to the entity that sells the API, the business tier and everything shipped next.
Evidence · Media corpus · 1,978 sources, mention-level parse, Sep 2025 – Mar 2026
Three self-descriptions run in parallel
Research lab, consumer product company, and infrastructure provider all appear in your own copy, sometimes on adjacent pages. Each is defensible. Running all three without a hierarchy means the reader picks one, and which one they pick decides whether they arrive expecting a paper, an app, or a contract.
Evidence · Site crawl · 1,040 pages, homepage and top-level section copy compared
Business profile
Four distinct businesses share one brand, one homepage and one word. Each has a different buyer, a different vocabulary and a different objection. The site treats them as sections rather than as separate audiences with separate journeys.
- Consumer subscription
- ChatGPT free and paid tiers, sold to individuals with no procurement step. The buyer arrives already knowing the product name and needs one thing from the page: what the paid tier does that the free one does not.
- Developer platform
- Models, SDKs and documentation sold by usage. This audience is served best of the four: the reference material is deep, the examples run, and pricing is stated in units a developer can forecast. It is also the only audience whose vocabulary you rank for consistently.
- Business and enterprise tiers
- Team and enterprise plans aimed at organisations that need administration, data controls and a contract. The public material describes features clearly and describes the security posture in far less depth, which inverts what a procurement reviewer actually needs first.
- Research and model releases
- System cards, model announcements and research posts function as the company’s most effective marketing asset. They generate the earned coverage that produces recall, and they are the pages least connected to any commercial next step.
Market position
You are the default answer to “which one should I use” and a contested answer to “why”. Anthropic argues safety and enterprise fit, Google DeepMind argues distribution, Meta and Mistral argue open weights and cost. The one argument nobody makes on your behalf is the one you would choose yourself.
Rivals define themselves against you, which sets the terms
Anthropic, Mistral and the open-weight community all position by contrast: safer, cheaper, sovereign, inspectable. Being the reference point is an advantage until the comparison axes are all chosen by other people. You currently rebut on capability and release cadence, which does not answer a question about control.
Evidence · Competitor site crawl · 5 rivals, positioning copy and comparison pages, Mar 2026
The strongest position you hold is unclaimed in copy
Across developer sources, the recurring reason to choose you is that the platform is the one everything else is written for – SDKs, tutorials, job posts and integrations assume it. That is a genuine moat and it appears nowhere as a stated claim on your own pages.
Evidence · Developer corpus · 14 communities, 3,100 mentions parsed, Sep 2025 – Mar 2026
Competitor landscape
On owned documentation depth you sit mid-table, not top. Two rivals publish more pages about their own products than you do, and the open-weight ecosystem publishes an order of magnitude more than anyone. Awareness is yours; the written record of the category is shared.
| Company | Public one-line read | Owned pages crawled | Top-3 holds, 24 queries | Assistant default framing |
|---|---|---|---|---|
| OpenAI | The company that made ChatGPT | 1,040 | 7 | Named first, described loosely |
| Anthropic | The safety-forward Claude company | 1,310 | 5 | Named second, described precisely |
| Google DeepMind | Research arm behind Gemini | 2,180 | 6 | Named with Google’s distribution |
| Meta | Publisher of open-weight Llama models | 1,470 | 3 | Named as the free option |
| Mistral | European open-weight challenger | 620 | 2 | Named on sovereignty questions |
| Open-weight ecosystems | Community hubs and forks | 9,400 | 9 | Named as the default for tinkering |
Anthropic out-documents you on the enterprise question
Their published material spends proportionally more space on data handling, deployment options and model behaviour than yours does. A reviewer comparing the two on public evidence alone finds more of what they were sent to find on the smaller company’s site. That is a content decision, not a product one.
Evidence · Competitor crawl · 1,310 pages, section-level classification, Mar 2026
Open-weight hubs win the queries that start a project
Community hubs hold nine of the 24 category top-three slots, mostly on how-to phrasing: running a model locally, comparing weights, fine-tuning on your own data. Those queries are the front door to a build decision, and you are not on the page where it gets made.
Evidence · Search · top-100 results for 24 category queries, Mar 2026
Brand analysis
The name is at ceiling and the meaning is not. Visual and verbal identity are disciplined on the marketing pages and loosen across documentation, help centre and research posts, where three different tones are doing the talking. Recognition this strong can carry an inconsistent message for a long time, and then it cannot.
Website analysis
openai.com is built for someone who already knows why they came. The four audiences are separated cleanly at the top level, and then the paths diverge across domains so that anyone comparing consumer and platform pricing has to hold two mental models at once. Nothing here is broken; several things are unhelpful.
Pricing lives in two places with two units
Consumer tiers are priced per seat per month; platform usage is priced per token. Both are clear on their own page and neither page acknowledges the other. A team lead deciding whether to buy seats or build on the API is the most common buyer in the middle, and there is no page written for them.
Evidence · Site crawl · consumer and platform pricing pages, cross-link check
Research posts are your best pages and your deadest ends
They earn the most external links in our sample and carry the fewest onward paths. A reader finishes a system card with maximum interest and maximum ambiguity about what to do next. One contextual next step per post would convert attention that you already paid for.
Evidence · Site crawl · research section, outbound-link and internal-link counts
118 pages ship without a description
Search engines and assistants both write their own summary when you decline to. On research and help pages that means your most technically precise material gets summarised by a machine reading the first paragraph. This is an afternoon of work with a measurable effect on how you are quoted.
Evidence · Site crawl · 1,040 pages, head-tag audit, Mar 2026
SEO analysis
You rank first for everything you named and lose the words that describe what you sell. Tutorial sites, newsrooms and encyclopedias hold the top three on 17 of 24 core category queries. The traffic you lose there matters less than the definitions you lose with it.
Your category vocabulary belongs to explainer sites
For queries like “what is a context window”, “fine-tuning vs prompting” and “how do embeddings work”, the pages ranking above you were written by tutorial publishers and news desks. They are competent, they are dated, and they are the first definition a new buyer reads. Documentation is not the same asset as an explainer written for someone who has not decided yet.
Evidence · Search · top-100 results for 24 category queries, Mar 2026
Branded queries are answered by aggregators before you
On several comparison-shaped queries containing your product names, the first result is a third-party review or a listicle. Those pages get your feature table wrong in ways nobody malicious intended. A maintained comparison page on your own domain would take that slot and fix the record.
Evidence · Search · 31 branded and comparison queries, top-10 results, Mar 2026
Documentation ranks and marketing does not
platform.openai.com holds top-three positions on 6 of the 7 category queries you win. The developer team is doing the SEO work for the whole company by accident, which is why the pages that rank are the ones written for people who already chose you.
Evidence · Search · domain-level position audit across 24 queries
Help centre content is invisible to search
A large share of help articles sit behind paths that rank for nothing and carry no descriptions. Every one of them answers a real question a prospect is typing into a search box. Publishing the top 40 as indexable pages is the cheapest ranking gain in this report.
Evidence · Crawl · help.openai.com, index and metadata check, Mar 2026
Digital presence
Owned channels punch below the brand. The developer forum and the public code repositories carry more useful signal than the corporate social accounts, and both are run as support surfaces rather than as places where the company explains itself. Reach is not the constraint here; intent is.
- Code repositories do the teaching
- The public cookbook and SDK repositories are where builders learn the product, and they are maintained well. They are also the least linked-to assets from the marketing site, which means the best teaching material you own is found by people who already went looking.
- Video is an event archive, not a library
- Conference talks and launch streams are published and then left as a chronological list. Cutting them into topic-titled segments would turn an archive into something a search engine and an assistant can quote from.
- The forum answers questions the site does not
- Recurring threads cover rate limits, model selection and migration between model versions. Three of those threads deserve to be permanent pages on your own domain, written once and kept current.
Customer perception
Sentiment splits cleanly by audience. Developers rate the platform on reliability and documentation and are broadly satisfied; consumer reviewers rate the product on usefulness and are enthusiastic; business buyers write about change management and data policy and are the least served group in the record. Nobody is unhappy about the product. Some people are unsure about the terms.
Model changes are the top recurring complaint
The most repeated theme across developer and business sources is not quality; it is change. People build a workflow, the behaviour of a model version shifts under them, and they cannot tell from your pages what changed. A plain-language change log written for non-engineers would remove the single most common source of frustration in this corpus.
Evidence · Reviews and forums · 4,180 reviews and 960 threads, theme frequency ranking
Documentation is praised by name
Reviewers who mention the developer documentation are overwhelmingly positive, and several name it as the reason they stayed after evaluating an alternative. That is rare and worth saying out loud on a page where an evaluator will read it.
Evidence · Reviews · 4,180 items, feature-mention sentiment split
Business buyers write about policy, not features
In the business-tier subset, the recurring words are retention, training, residency and approval. Your product pages answer with capability lists. The mismatch is not a trust deficit; it is a topic mismatch between what they ask and what the page discusses.
Evidence · Reviews · business-tier subset, 640 items, keyword frequency
Trust signals
The trust material is credible, specific and hard to find. Usage policies, model behaviour documentation and enterprise privacy commitments all exist and read like they were written by people who meant them. There are 23 such pages against 41 product and research pages, and the shortest path from the homepage runs three clicks.
Your safety documentation is a sales asset filed as a compliance one
System cards and behaviour specifications describe how models are evaluated in more detail than most buyers expect to get. Presented as evidence of diligence rather than as technical appendices, they would answer the enterprise objection before it is raised.
Evidence · Site crawl · safety and policy sections, 23 pages classified
Trust content sits below the fold of the buyer journey
No path from the business tier page reaches security material in fewer than three clicks, and none of the pricing pages link to it at all. Buyers who need that information will find it; buyers who needed to be reassured before they asked will not.
Evidence · Site crawl · path analysis from homepage and business pages
Third parties are writing your policy summary
For queries about data handling and training practices, the top results are news articles and analyst posts rather than your own policy pages. Those summaries are mostly accurate and always partial, and they are what a reviewer reads first.
Evidence · Search · 12 policy-shaped queries, top-10 results, Mar 2026
- What is present
- Usage policies, model behaviour documentation, enterprise privacy commitments, system cards for major releases, and a security contact route. The substance is there.
- What is thin
- A single consolidated page a legal reviewer can send to a colleague. Right now that person assembles one from five sources and forwards a folder of links.
- What is missing
- Plain-language answers to the four questions that appear most in business reviews: what is retained, what is used for training, where it runs, and who approves access.
Technology stack
The developer surface is the strongest thing the company publishes. Official SDKs, a machine-readable API specification, streaming, tool calling and embeddings are documented to a standard that competitors copy. The weakness is that this knowledge lives across four domains and a code host, with no single map of where anything is.
- Documented client libraries
- First-party SDKs in the two languages most builders reach for, with community libraries covering the rest. Version notes are current and the examples run without editing, which is a lower bar than it should be and one most rivals miss.
- A machine-readable API specification
- Publishing the specification lets other tools generate clients and lets assistants answer questions about your API correctly. No other page you publish does as much work per word, and it is not mentioned anywhere a non-developer would see.
- Reference implementations in a public cookbook
- Working examples for retrieval, tool use, evaluation and structured output. This is the material that turns an evaluation into a build, and it is on a domain most buyers never visit.
- Four documentation homes, no map
- Marketing, platform documentation, help centre and cookbook each answer part of a question. There is no page that says which one to read for what, so people search and arrive wherever the index sent them.
AI visibility
Every assistant we asked names you first, and half of them describe you using someone else’s words. That is the whole finding: presence is guaranteed, framing is borrowed. Assistants quote the pages that explain a category clearly, and on your category the clearest pages are not yours.
Citations go to encyclopedias and news before they go to you
When an assistant explains what the company does, the supporting link is more often a reference site or a newsroom than openai.com. Those sources update slowly. Anything you shipped recently is described by a page written before it existed.
Evidence · Assistant runs · 48 prompts across four assistants, citation domains logged, Mar 2026
The developer answer is accurate everywhere
Put an integration question to any assistant in the set and the reply comes back correct, current and sourced to your documentation. The pattern is clean: where you publish the clearest page, the assistants use it. That is the argument for writing the buyer-side explainers you have not written.
Evidence · Assistant runs · 16 developer-intent prompts, answer accuracy and citation check
Names the company first and describes it through its own product line, citing owned documentation on technical follow-ups.
Names the company first and gives a precise, current description, then presents Anthropic and Google DeepMind as direct alternatives in the same breath.
Names the company first but sources the description from news coverage, so recent releases are summarised from articles rather than from your pages.
Names the company first and cites encyclopedia and third-party explainer pages ahead of openai.com, which dates the description by several months.
Growth opportunities
Upside that is reachable with the business you already have – ordered by how much of it is within your own control.
Own the words your category runs on
Seventeen of 24 core queries are held by explainer sites. Twelve original explainers – written for a buyer, not a builder – would put your definitions in front of people at the moment they form an opinion. You already have the subject-matter depth; what is missing is a page addressed to someone who has not decided yet.
Evidence · Search · top-100 results for 24 category queries, Mar 2026
Publish the help centre as indexable pages
A large share of help articles rank for nothing and answer questions prospects are actively typing. Reworking the top 40 into public, described, linkable pages is days of work and moves both search position and assistant citation on the same content you already wrote.
Evidence · Crawl · help.openai.com index and metadata audit
Turn research posts into an on-ramp
Research pages earn the most external links in our sample and offer the fewest next steps. One contextual path per post, matched to the topic rather than a generic call to action, converts attention that is already arriving and already paid for.
Evidence · Site crawl · research section, inbound-link and internal-link counts
Say the thing developers already say about you
The most repeated reason to choose the platform is that everything else is written for it. That claim appears in 3,100 community mentions and on none of your own pages. Stated plainly on the platform page, it is the strongest sentence you are not using.
Evidence · Developer corpus · 14 communities, 3,100 mentions parsed
Risks
Structural exposure visible in the public record. None of these are predictions; each one is something already observable.
Rivals are choosing the comparison axes
Safety, cost, openness and sovereignty are the four dimensions the category is now argued on, and all four were introduced by someone else. Answering on capability does not address a question about control. Left alone, the comparison table a buyer builds will be a competitor’s table.
Evidence · Competitor crawl · 5 rivals, positioning and comparison pages, Mar 2026
The enterprise reviewer leaves without an answer
Twenty-three trust pages against 41 product pages, three clicks from the homepage, none linked from pricing. The reviewer either assembles the answer from news coverage or asks a salesperson, and the first of those two paths happens more often.
Evidence · Site crawl · path analysis and section-level page classification
Model change is the loudest complaint in the record
Across 4,180 reviews and 960 threads, the top recurring theme is behaviour shifting under a built workflow with no plain-language account of what changed. This is a documentation gap that reads to customers as an operational one.
Evidence · Reviews and forums · theme frequency ranking, Jan 2025 – Mar 2026
Assistants are describing you from stale sources
Just over half of the citations in our assistant runs point at domains you do not control, many of them reference pages that update in months rather than days. Your release cadence is faster than the pages being quoted about it.
Evidence · Assistant runs · 48 prompts, citation domain and freshness logging
Strategic recommendations
What we would do, in order, if this were our own position to defend.
Write one sentence and enforce it everywhere
Pick the single description the company leads with and put it at the top of the homepage, the platform page, the business page and the research index. Five competing self-descriptions is the root cause of most findings in this report, and it is fixed by a decision rather than by a project.
Evidence · Site crawl · homepage and section-level positioning copy
Build twelve buyer-side explainers
One page per category term where an explainer site currently holds the top three. Written for the person evaluating, not the person implementing, each ending in the decision it supports. Target the queries with commercial intent first.
Evidence · Search · 24 category queries, intent classification and current holders
Consolidate trust into one sendable page
One URL a legal reviewer can forward, covering retention, training use, deployment location and access approval in plain language, linked from pricing and the business tier. The material exists across five pages; this is assembly, not authorship.
Evidence · Site crawl · 23 trust pages, overlap and gap analysis
Publish a change log written for non-engineers
Dated entries describing what changed in observable behaviour and what a customer should check. It answers the most repeated complaint in the review corpus and gives assistants a current page to cite when someone asks what is new.
Evidence · Reviews and forums · top recurring theme, 4,180 reviews and 960 threads
Give the documentation estate one front door
A single index that says which of the four domains answers which question, linked from the main navigation. Cheap to build, and it repairs the specific failure where people search for something you documented and land on a page written by somebody else.
Evidence · Crawl · four owned documentation domains, cross-link analysis
90-day action plan
Three windows. The first pays for the report, the second compounds it, the third only works once the first two are done.
Fix 118 missing descriptions
Write meta descriptions for every page shipping without one, prioritising research posts and help articles. This decides how search engines and assistants summarise you, and it is the fastest measurable change available.
Choose the one-sentence description
Settle the research lab, consumer product and infrastructure question internally and publish the winner at the top of four key pages. Everything downstream in this plan depends on which sentence you pick.
Ship the consolidated trust page
One URL covering retention, training use, deployment location and access approval, linked from pricing and the business tier. Assemble from existing policy material rather than writing new commitments.
Publish the first six explainers
Start with the category terms where a tutorial site holds all three top slots and the query has purchase intent. One page per term, written for an evaluator, each closing with the decision it supports.
Open the help centre to search
Convert the top 40 help articles into indexable pages with descriptions and internal links from the relevant product sections. Measure position change on the questions they answer at day 90.
Build the documentation front door
One index page mapping the four documentation domains to the questions each answers, linked from primary navigation and from the cookbook. Re-run the assistant citation check afterwards to confirm the shift.
Intelligence Score
You own the category’s attention but not its definition – the explaining is being done on other people’s pages.
Company Intelligence Score
86/100
The headline score is the plain mean of the six dimensions below. No weighting, no curve.
Near-total unprompted recall, held back only by how differently people finish the sentence “OpenAI is the company that…”.
You rank first for your product names and lose the top three to tutorial sites and newsrooms on 17 of 24 category-vocabulary queries.
Every assistant we queried names you first in the category, but two of four describe you through third-party sources rather than your own pages.
Security, privacy and model-behaviour material exists and is credible, but it is thin against the product estate and buried three clicks from the homepage.
Default choice in consumer and developer conversation, with the enterprise story argued about rather than owned.
Distribution, developer habit and product release cadence all point the same direction; the constraint is narrative, not demand.
Keep reading
Three more, same rubric.
The rubric only means something across subjects. Read a second one and the shape of the first starts to explain itself.
Founder Intelligence
Alexandra Reyes
Real credentials, no owned surface, invisible to assistants.
Illustrative demonstration, not a real assessment.
Automotive & energy
Tesla
Category-defining brand, thinning narrative control.
Illustrative demonstration, not a real assessment.
Financial infrastructure
Stripe
Developer trust at ceiling, buyer-side story unfinished.
Illustrative demonstration, not a real assessment.
Now run it on your own company.
Same format, same rubric, more sections than this one – applied to your own public record. About six minutes from here.
- One-time payment – no subscription
- Readable for 90 days
- Export to PDF, yours to keep
- Secure processing via Stripe