Do Investor Materials Get Cited by AI?

What “cited” has to mean here

The question in the title cannot be answered the way people want it answered, and it is worth being precise about why before producing numbers.

No assistant operator publishes citation rates by source type. Microsoft shows site owners their own citation counts and warns that the figure indicates nothing about “ranking, authority, or the role of any page within an individual answer”; Google publishes eligibility conditions and not selection; OpenAI publishes crawler controls and not selection. That is the whole state of the art, set out in what signals answer engines use. The independent measurement that does exist — the Stanford-affiliated evaluation covered in which sources AI assistants cite — is an evaluation of chatbots as news intermediaries, so it has nothing to say about filings by design.

So we cannot tell you a percentage. What we can do is measure the material conditions that determine whether a document is available to be cited, and compare investor materials against press coverage on each of them: is it permitted to crawlers, is it actually crawled, does it persist, is it machine-readable, and is there any provenance attached. On all five, filings behave unlike almost anything else on the web, and that is the finding.

This site's own newsroom work keeps arriving at the same place from the other direction. When a claim about a company needs checking — revenue, headcount, a segment number, who owns what — the filing is routinely the only source that is dated, addressable, specific and attributable to a person who can be prosecuted for getting it wrong. The guide below is an attempt to say exactly why that is true in infrastructure terms rather than as a preference.

Is EDGAR crawled? Yes, and the SEC asks for it

Start with permission, because everything downstream depends on it. The SEC's robots.txt is a long file that mostly reflects the content-management system underneath the website, and then a short block at the end headed #SEC does the real work. The first line of it is Allow: /Archives/edgar/data. The lines after it disallow /cgi-bin, /Archives/bin, /Archives/etc, /Archives/usr and a small set of redaction paths.

Read that as an instruction rather than a formality: the documents are explicitly open and the interactive query screens are explicitly closed. A crawler is being told to take the filings and leave the search forms alone. Almost no commercial publisher's robots.txt reads like that, and several of the ones examined in do news sites block AI crawlers read like the opposite.

Crawlers accept. We queried the Common Crawl index for the September 2026 crawl, CC-MAIN-2026-39, with a single prefix query on sec.gov/Archives/edgar/data/. It returned 29,766 records. The fetch timestamps in those records run from 4 September to 17 September 2026, which is the crawl's own window. The URLs cover 7,163 distinct company identifiers — the numeric CIK that follows /data/ in every EDGAR path — so this is breadth across the registrant population, not depth on a handful of famous issuers. By content type, 23,827 were HTML, 1,723 were XML and 944 were plain text, which is roughly the shape of EDGAR itself: exhibits and primary documents in HTML, XBRL instance and schema files in XML, and older submissions in text.

The status distribution is the part worth keeping. 26,518 returned HTTP 200. 1,680 were 301 redirects, mostly the host and casing normalisations EDGAR has accumulated over thirty years. 1,183 were 304s recorded as warc/revisit, meaning the crawler had the content already and the server confirmed it had not changed. 224 were 503s, which is what a rate-limited crawler gets. And 155 — half of one per cent — were 404.

Half a per cent. For comparison, in the same crawl we found a 0.2% 404 rate on PR Newswire's own release pages and a 45.1% rate on one partner site's copies of those same releases, which is the subject of where a wire release actually ends up. Publisher-side decay is worse again: Pew's sample found 16% of pages individually inaccessible on root domains that were otherwise working, and the author-page and masthead cases are in what happens to coverage when a journalist leaves. A government filing archive with a statutory retention obligation is simply a different class of object from a news URL.

What the assistants' own documentation says: nothing

Having established that the material is open and crawled, the honest next step is to check whether anyone building an assistant says they use it. We read the three published crawler documents that matter on 26 September 2026.

OpenAI documents four crawlers. OAI-SearchBot is “used to surface websites in search results in ChatGPT's search features”. GPTBot is “used to make our generative AI foundation models more useful and safe”. ChatGPT-User is “used for certain user actions in ChatGPT and Custom GPTs”. OAI-AdsBot is “used to validate the safety of web pages submitted as ads on ChatGPT”.

Anthropic documents three. ClaudeBot “helps enhance the utility and safety of our generative AI models by collecting web content that could potentially contribute to their training”. Claude-User “supports Claude AI users” when they ask questions. Claude-SearchBot “navigates the web to improve search result quality for users”.

Google lists eleven common crawlers, from Googlebot, “used to find information for building Google's search indexes”, through the image, video and news variants, to Google-Extended, which lets site owners “manage whether content Google crawls from their sites may be used for training future generations of Gemini models”.

Across all three documents, the number of references to EDGAR, the SEC, regulatory filings, investor relations or any named financial data source is zero. None of them explains how a source is selected for a citation either, which is consistent with what every other operator does and does not publish.

This is where a guide like this one has to stop and refuse to fill the gap. The infrastructure argument below is strong. It is not a citation measurement, nobody has published one for filings, and the correct posture is that filings are unusually well positioned to be used and that no operator has confirmed it. If you read a confident claim that a particular assistant “prefers SEC filings”, ask who measured it and against what.

The part nobody exploits: filings are a database

The strongest argument for investor materials is not that they are trustworthy prose. It is that they are not prose at all.

The SEC publishes RESTful JSON APIs on data.sec.gov with no authentication. Its documentation describes four: a submissions endpoint giving each entity's filing history with names, exchanges and tickers; a company-concept endpoint that “returns all the XBRL disclosures from a single company (CIK) and concept (a taxonomy and tag) into a single JSON file”; a company-facts endpoint that “returns all the company concepts data for a company into a single API call”; and a frames endpoint which “aggregates one fact for each reporting entity that is last filed that most closely fits the calendrical period requested”. The SEC states they are “updated in real-time as filings are disseminated”, with a typical processing delay “of less than a second” for submissions and “under a minute” for the XBRL APIs.

We used two of them. Asking the company-concept endpoint for Apple's us-gaap:Assets returned 146 individual facts in a 20KB file, the most recent being $383,266,000,000 as at 27 June 2026, tagged to accession number 0000320193-26-000020, form type 10-Q, filed 31 July 2026. Every number arrives with the document it came from and the date it was filed. Then we called the frames endpoint for us-gaap:Revenues in US dollars for calendar 2025 and got back a single 343,855-byte file containing 2,243 companies, each with a CIK, a legal entity name, a value and a period end date. One HTTP request, no key, no scraping, 2,243 audited revenue figures.

Compare that with the equivalent question asked of press coverage. A revenue figure in an article is a string inside a sentence, unlabelled, undated relative to the fiscal period, unattributed to a source document, and subject to the corrections behaviour described in how publications handle corrections. A revenue figure in XBRL is typed, scoped to a period, bound to a taxonomy concept and traceable to an accession number. Whether general-purpose markup on a web page helps at all is a separate and much weaker question, examined in does schema markup affect citation. Filings are not marked-up prose; they are a database with a document attached.

And the structured layer is itself being crawled. Our query on data.sec.gov returned 365 records, 362 of them HTTP 200, and 357 of those carried the MIME type application/json. Crawlers are fetching the API responses, not only the filing documents.

Provenance: the thing a press release cannot buy

There is one property of a filing that no amount of distribution can replicate, and it is easy to overlook because it lives in the exhibits nobody reads.

The annual report we worked through for this research — ACCESS Newswire's 10-K for 2025, filed 19 March 2026 — carries four certification exhibits. Exhibit 31.1, under Rule 13a-14(a) and Section 302 of the Sarbanes-Oxley Act, begins “I, Brian R. Balbirnie, certify that: 1. I have reviewed this Annual Report on Form 10-K of ACCESS Newswire Inc.; 2. Based on my knowledge, this report does not contain any untrue statement of a material fact or omit to state a material fact necessary to make the statements made … not misleading”. Exhibit 32.1, under 18 U.S.C. Section 1350 and Section 906, is signed by the same named chief executive and certifies that the report “fairly presents, in all material respects, the financial condition and results of operations of the Company”.

A named human being has signed, in a public document, a statement that the contents are true, under a statute with criminal penalties. That is a provenance chain. A press release has an issuing company, a distributor that checks it against editorial guidelines rather than facts — as will a wire publish anything sets out — and no certification whatever. An article has a byline and a corrections policy, which is meaningful but weaker, and which decays in the ways how long a placement stays findable describes.

Whether a retrieval system explicitly weighs any of this is unknown. But if you are deciding where to put a number that you want a machine to repeat accurately in two years, the filing is the version with a signature on it.

The investor relations site is the weak link

Given all of the above, the surprise is how little of it the companies themselves build on. The IR website is the one part of the investor stack a company controls completely, and it is the part doing least.

We fetched seven IR homepages. Apple's returned 200 and contained zero JSON-LD blocks. Alphabet's returned 200 and contained zero. Microsoft's contained three: two Organization objects with sameAs links to its social profiles, and an FAQPage whose first question is “Know the next earnings release date”. Microsoft is also the only one of the three running its investor section on its own web platform. Tesla's, Nvidia's, Coinbase's and Salesforce's IR sites returned HTTP 403 to our client, and ir.tesla.com/robots.txt returned an Akamai “Access Denied” page, so we cannot describe what they carry.

The robots files told a story we did not expect. investor.apple.com/robots.txt, abc.xyz/robots.txt, investor.nvidia.com/robots.txt, investor.coinbase.com/robots.txt and investor.salesforce.com/robots.txt are the same file: 207 bytes each, identical checksums, five unrelated companies. The contents are a wildcard Crawl-delay: 10 with Allow: /, followed by named exceptions with a 60-second delay for Eluta.ca, MS Search 6.0 Robot and gsa-crawler — a Canadian job-search spider, a Microsoft SharePoint crawler and Google's on-premises search appliance, none of which has been a meaningful crawler for a decade.

Two things follow. First, the IR estate of the largest companies in the world is a template estate run by a handful of vendors, and the default configuration is old enough to be addressing software that no longer exists. Second, and more usefully: no AI crawler is named in any of those files, so under the robots exclusion standard every one of them falls through to Allow: /. Corporate IR sites are wide open by omission, in exactly the way which AI crawlers publishers name found news sites are not.

So the practical advice for a public company is the reverse of the usual: your filings are already in excellent shape and your IR site is a vendor default. Marking up the IR newsroom, the officer biographies and the event calendar costs almost nothing and is the only part of this that is yours to fix. The entity-level version of that argument is in how an entity enters the knowledge graph.

If you have no filings

Most companies reading this are private, and the temptation is to conclude that none of it applies. That is half wrong.

You may already be on EDGAR. On 25 September 2026 the daily index shows 147 Form D filings and 155 Form D/A amendments — 302 of the day's 4,411 filings — and Form D is the notice of an exempt offering under Regulation D, which is what most venture-backed private companies file when they raise a priced round. A Form D is thin: entity name, jurisdiction, address, year of incorporation, industry, executive officers and directors by name, and the size of the offering. It is not audited and it is not certified the way a 10-K is. It is also a permanent, government-hosted, machine-readable record that names your company and your officers, and for a private company it is usually the only one that exists. Check whether yours is there, check that it is right, and cite it when you need a primary source for your own corporate facts. Reg A issuers have more: the same index shows a 1-A POS filing that day.

Everything else has to be built. Absent filings, the properties that make investor materials citable have to be reconstructed from scratch: a permanent URL, a dated statement, a named accountable person, and machine-readable structure. That means an on-domain facts page with Organization markup that actually validates, officer pages with Person markup, a dated newsroom, and identifiers that let a machine tie the records together — the subject of do identifier registries affect search visibility. It is a poorer substitute than a filing and it is considerably better than a press release.

What it is not is a reason to buy more coverage. The failure mode we see most often is a private company treating a funding announcement on a wire as though it were a filing. It is not the same object: no certification, no structure, no permanence for the syndicated copies, and a link profile that passes nothing. Where coverage genuinely does work is a separate question with its own evidence, and paywalled coverage has its own indexing quirks, covered in does paywalled coverage get indexed.

Methodology and limitations

Everything here was read or queried on 26 September 2026. Primary sources: sec.gov/robots.txt; the SEC's “Accessing EDGAR Data” page and its EDGAR API documentation; the data.sec.gov company-concept endpoint for CIK 0000320193 (us-gaap:Assets) and the frames endpoint for us-gaap:Revenues/USD/CY2025; the EDGAR daily index by form type for 25 September 2026; EDGAR company search for CIK 843006 and the resulting 10-K and certification exhibits 31.1 and 32.1; the Common Crawl index API for collection CC-MAIN-2026-39; OpenAI's, Anthropic's and Google's published crawler documentation; and the robots.txt and homepage HTML of seven investor relations sites. Common Crawl queries used a single request per path with matchType=prefix, output=json and pageSize=1000, counting records and their status codes. Seven limitations. First and most important, nothing here measures citation. We measured availability, persistence, structure and provenance; whether any assistant weighs those properties is not published by any operator and we have not inferred a rate. Second, Common Crawl record totals are not a census: repeated queries with different pagination parameters return different totals, so we report status shares within one result set and state the sample size each time. Third, Common Crawl is one crawler with one seeding strategy; a 200 in its log is evidence that a well-behaved anonymous crawler could fetch the page, not that any particular company's crawler did. Fourth, the IR sample is seven homepages, four of which refused us, so the structured-data finding is an illustration of a pattern rather than a survey, and inner pages may carry markup the homepage does not. Fifth, the certification language is quoted from one issuer's exhibits; the wording is largely statutory and highly standardised across filers, but we verified it on one filing, not many. Sixth, the Form D argument concerns the existence and durability of the record, not its quality: a Form D is unaudited, often minimal, and sometimes stale. Seventh, EDGAR's reliability is a property of a US regulator and does not transfer automatically to other jurisdictions' filing systems, several of which we did not test.

Conclusion

The honest answer to the title question is that investor materials are better positioned to be cited than almost any other corporate source, and that no assistant operator has confirmed it. The positioning is not a matter of opinion: the SEC's robots.txt says Allow: /Archives/edgar/data; a single September 2026 crawl query returned 29,766 EDGAR URLs across 7,163 companies with a 0.5% loss rate, against 45.1% for one partner site's copies of press releases in the same crawl; the SEC publishes unauthenticated JSON APIs updated “in real-time as filings are disseminated”, one call to which returned 2,243 companies' 2025 revenue; and every filing carries a named officer's certification, under a criminal statute, that it contains no untrue statement of a material fact. Nothing in public relations has any of those four properties. Against that, OpenAI's, Anthropic's and Google's crawler documentation mentions the SEC exactly zero times, and no operator publishes how sources are chosen, so the causal claim stays unproven and should be described that way. The practical reading is straightforward. If you have filing obligations, your filings are already the strongest machine-readable record you own, and the thing worth fixing is the investor relations site — which, on the evidence of five identical 207-byte robots files, is currently a vendor default optimised for crawlers that stopped existing a decade ago. If you have no filings, check whether a Form D already names you, and build the properties a filing has rather than buying more of the properties a press release has.

Frequently Asked Questions

Sources

1. documentation U.S. Securities and Exchange Commission, robots.txt. The #SEC block, including Allow: /Archives/edgar/data and the /cgi-bin disallow. Read 26 September 2026.
2. documentation U.S. Securities and Exchange Commission, Accessing EDGAR Data. “Anyone can access and download this information for free”; “Current max request rate: 10 requests/second”; the user-agent declaration request; bulk feed archives. Read 26 September 2026.
3. documentation U.S. Securities and Exchange Commission, EDGAR Application Programming Interfaces. The submissions, company-concept, company-facts and frames endpoints; “updated in real-time as filings are disseminated”; processing delays. Read 26 September 2026.
4. live read U.S. Securities and Exchange Commission, frames API, us-gaap:Revenues, USD, CY2025. 2,243 reporting entities in one 343,855-byte response. And the company-concept response for CIK 0000320193: 146 facts, most recent $383,266,000,000 at 27 June 2026 from accession 0000320193-26-000020, Form 10-Q filed 31 July 2026. Retrieved 26 September 2026.
5. live read U.S. Securities and Exchange Commission, EDGAR daily index by form type, 25 September 2026. 4,411 filings, 120 form types, 147 Form D, 155 Form D/A, 207 combined 10-K/10-Q/8-K. Read 26 September 2026.
6. filing ACCESS Newswire Inc., Exhibit 31.1, certification under Rule 13a-14(a) and Section 302 of the Sarbanes-Oxley Act and Exhibit 32.1, certification under 18 U.S.C. Section 1350, both filed with the Form 10-K for the year ended 31 December 2025 on 19 March 2026. Read 26 September 2026.
7. operator Common Crawl, index API, collection CC-MAIN-2026-39. Prefix query on sec.gov/Archives/edgar/data/ returned 29,766 records (26,518 × 200, 1,680 × 301, 1,183 × 304, 224 × 503, 155 × 404, 3 × 303, 3 × 403) across 7,163 distinct CIKs, fetched 4–17 September 2026; prefix query on data.sec.gov returned 365 records, 362 × 200, 357 application/json. Queried 26 September 2026.
8. documentation OpenAI, OpenAI crawlers. OAI-SearchBot, OAI-AdsBot, GPTBot and ChatGPT-User with the quoted purposes. No mention of EDGAR, the SEC or financial sources. Read 26 September 2026.
9. documentation Anthropic, Does Anthropic crawl data from the web, and how can site owners block the crawler? ClaudeBot, Claude-User and Claude-SearchBot with the quoted purposes. No mention of EDGAR, the SEC or financial sources. Read 26 September 2026.
10. documentation Google Search Central, Google crawlers and fetchers. Eleven common crawlers including Googlebot and Google-Extended. No mention of EDGAR, the SEC or financial sources. Read 26 September 2026.
11. vendor platform Investor relations sites, read 26 September 2026: investor.apple.com/robots.txt, abc.xyz/robots.txt, investor.nvidia.com/robots.txt, investor.coinbase.com/robots.txt and investor.salesforce.com/robots.txt — all 207 bytes with identical MD5 checksums. Homepage JSON-LD: Apple 0 blocks, Alphabet 0 blocks, Microsoft 3 blocks. Tesla, Nvidia, Coinbase and Salesforce IR pages returned HTTP 403 to our client.
12. peer-reviewed Pew Research Center, May 2024. When Online Content Disappears. Cited here only for the 16% of sampled pages individually inaccessible on root domains that still worked, as the comparison case for filing persistence. Read 26 September 2026.