How we work, and what you should know about us

Sources: filings first

Every number in our research comes from what companies file with the US Securities and Exchange Commission: the annual 10-K, the quarterly 10-Q, earnings releases furnished on Form 8-K, and the annual 20-F for foreign issuers. Financial figures are read from the structured XBRL data attached to those filings, including segment disclosures, so they match what the company reported to the regulator rather than a summary in the press. Where we describe a relationship between companies, a lease commitment or a customer concentration, we quote the filing’s own sentence and link to the document.

We don’t use analyst estimates, consensus forecasts, management guidance as fact, or paid data feeds in our published research. When something isn’t disclosed, we say so rather than estimate it. Where a page depends on an assumption that no filing provides, such as the share of capital spending that goes to servers, the page labels it as an assumption and lets you change it.

How figures are checked

The data is pulled from SEC filings by scripts that record the filing behind every figure. Quarterly values are taken directly where companies report them and derived from year-to-date figures where they don’t; companies with different fiscal years are aligned to the nearest calendar quarter before anything is added up, and the pages say so. A financing link or disclosure is published only if the filing’s own sentence contains the figures we cite; if a later filing drops it, it drops off the site rather than going stale.

Text that cites a figure is generated from the same data as the charts beside it, so a refresh can’t leave a sentence saying one number and a chart showing another. Qualitative judgments, such as which layer kept the most profit, are re-read whenever the data is refreshed.

Errors we design against

Financial data from filings has well-known traps, and most published errors come from a handful of them. Companies change the XBRL tag they use for the same line item from one year to the next, so our pull pools equivalent tags and only joins them when they agree on the periods both report. Many companies report cash flows year to date, so a quarter has to be derived by subtraction, and a fourth quarter only exists as the full year minus nine months. Fiscal calendars differ by up to two months, so figures are aligned to calendar quarters before they are added up. Companies restate past periods, so the latest filing wins for each period.

Text extraction has its own traps. Financial tables run into the sentences after them, headings run into the sentences that follow, and risk factors describe things that might happen in language that looks like disclosure of things that did. Our readers skip conditional sentences, keep only sentences that contain the figures we cite, and record the filing for every quote so anyone can check it.

What we choose to cover

Companies appear on the AI stack map when they file with the SEC and AI is a material part of their revenue or costs. We cover the companies that best show how money moves through each layer, not every company that mentions AI, and inclusion is never a view on the stock. Companies with less than $100 million of revenue appear in their layer’s table but stay out of totals and growth rankings, where a tiny base would distort the comparison. Private companies such as OpenAI and Anthropic appear only where a public company’s filing discloses a relationship with them.

We decide what to publish, and in what order, on one test: whether the filings support a page of real substance. That is why company pages started with twenty companies, then added enough for at least three in every layer, rather than every name on the map, and why new pages are added as the data allows rather than to fill a list.

Dated reports don’t change

The live research pages update whenever the data is refreshed, normally after each earnings season. The quarterly AI Stack Report is different: each report’s figures are frozen when it is published, so anyone citing it can rely on it still saying what it said. If a company restates a quarter in a way that would change a report’s conclusions, we add a dated note to the report instead of altering it.

Independence

We don’t accept payment to cover, feature, rate or omit a company, and companies don’t review our analysis before publication. We don’t publish sponsored research. Paid work (talks, board and investment-committee briefings, custom research and data licensing, described on the work-with-us page) is done for clients and kept separate from what we publish; a client can’t buy a change to a public page.

The AI Signal Gauge’s demonstration ratings use fictional companies with tickers that belong to no real listing, so a demonstration is never mistaken for a view on a real stock. Ratings of real companies are kept in a review sandbox while the method is calibrated and are not published.

Not investment advice

Nothing on this site recommends buying or selling a security or takes account of anyone’s circumstances. Our research explains how money moves through the AI industry and what companies disclose. Companies are named because they show how the stack works, not as a view on their shares. Decisions stay with you and, where appropriate, your adviser.

Corrections

We correct errors of fact promptly. A corrected figure on a live page is fixed at the next deploy with a note on the page saying what changed and when, unless the change comes from routine data refreshes, which the pages describe. Corrections to a dated report are added as dated notes on that report. To report an error, use the contact form with the page and the figure, or reply to any newsletter.

No corrections have been issued so far. Future corrections will be listed here with their date and the page they affect.

Disclosures

Investing in AI is run by Rob May and Vik Chadha. Both work in AI and invest in it, which is what qualifies them to write about it and also why readers should know their roles. These are their current positions outside this site:

Holdings

Our policy is to disclose any position either of us holds in a company we analyze. We are compiling the list of current positions in companies covered on this site and will publish it here. Until then, assume that we may hold shares in large AI companies, including companies covered on the AI stack map.

Commercial relationships

Any commercial relationship with a company we cover, or with a publisher or platform whose products we discuss, will be listed here. The AI Investor Assessment includes Rob May’s own book and newsletter in its results, and the assessment page says so.

Rob and Vik also co-own My Finance Platform, a separate personal-finance software business. It has no commercial relationship with any company covered here or with any publisher whose products we discuss. Neither of us owns shares in MarketWise, the parent of Chaikin Analytics, whose products appear in the assessment’s results.

Use of AI tools

We use software, including AI tools, to read filings, extract figures, check them against each other and draft text. Covering hundreds of filings a quarter would not be possible otherwise. The rules above apply to everything that is published, however it was produced: figures come from filings, quoted links must contain their figures in the filing’s own words, and Rob and Vik are responsible for what appears on the site.

Accountability

Rob May and Vik Chadha are responsible for everything published on Investing in AI. Research pages carry the date the underlying data was last pulled from SEC filings, dated reports carry their publication date, and every quoted disclosure carries the filing it came from, so a reader can always tell how current a figure is and where it came from. When we change a method in a way that moves published numbers, such as how a layer total is defined or which companies are included, the page that uses the method says so.

We welcome challenges to our reading of a filing. If you think a figure is wrong, or a quote is taken out of context, tell us through the contact form with the page and the filing, and we will check it and reply. Companies we write about can do the same; a correction is made when the filing supports it, not because a company asks. Readers can expect the same answer to the same question whoever is asking it, and corrections are made in public on the page concerned rather than quietly. That applies to our own mistakes in method as much as to individual figures.