From OSINT to Search: How to Investigate Public Information in the AI Era

The mosque was an important clue. Eliot Higgins was checking footage said to show fighters in Tiji, Libya, and used the road beside the mosque to narrow down its location. He compared the minaret and the curve of the street with satellite images, describing the process in a 2014 Bellingcat guide.

Open source intelligence (OSINT) uses public information, but access to a source is only the beginning. Search results can repeat a misleading caption, and an AI summary can repeat it again. Higgins checked the location against details another person could examine on a map.

What Are You Trying to Prove?

The first step in researching the Libya footage was deciding which claim needed checking. Higgins could match the location visible in the video with features on a map. That answered where the camera was pointed. It didn’t identify everyone in the footage or prove what happened before the recording began.

Business research needs the same precision. A company registry can confirm the registered name of a business. Its website can show which services it advertised when the page was checked. Neither record proves a particular customer received the service. If you don’t separate those questions, a folder full of documents can leave you with a conclusion none of them supports.

Be careful with the verbs in your notes. Listed means you saw the name or address in a source. Reported means somebody made a claim. Verified should be reserved for the specific point the available evidence supports. That language matters when an observation ends up in a report someone else will rely on.

Ask what record could settle the question and what would make you reconsider. It saves time, and it helps you recognize when you haven’t found an answer yet.

Build a Source Ledger Before the Investigation Gets Large

If a source matters to the answer, save enough information to find and interpret it again. A browser tab won’t preserve a page that later changes, and a screenshot without the URL or capture date can lose the context that made it useful. For a small investigation, an ordinary spreadsheet or structured note is enough.

• Reference and location. Give each source an ID and save its URL, archive capture, record number or API endpoint.
• Publisher and source type. Identify who controls the record and whether it’s an official statement, filing, independent report, directory or another kind of material.
• Dates. Keep the publication or filing date when available and the date you inspected the record. They answer different questions.
• Supported claim. Write exactly what the source establishes, with a relevant excerpt or field when needed.
• Limitations and conflicts. Note uncertain identity matches, missing context and records that disagree.
• Retrieval details. Retain important query parameters, search terms or tool outputs so someone else can repeat the check.

For work with legal or human rights consequences, evidence handling needs greater care. The Berkeley Protocol on Digital Open Source Investigations, developed by UC Berkeley’s Human Rights Center and the UN Human Rights Office, covers collection, preservation, verification, analysis and protection of people involved. Everyday business research doesn’t require the same process, but it benefits from careful records.

Choose Sources for the Claim You’re Testing

There isn’t a single ranking of sources that works for every investigation. A business registry can establish a registered company name, while a current service page tells you what the business says it offers. Neither necessarily answers whether a customer received the service or how it performed.

Choose records according to the fact you’re checking. Some common starting points are useful.

1. Original records and direct measurements for facts that can be established through filings, documents, public datasets or reproducible tests.
2. Current official statements for what an organization says about its own products, policies and operations.
3. Independent reporting when the reporter adds interviews, documents or observations that can be checked.
4. Directories, community posts and specialist databases for additional context and leads, with attention to who supplied the underlying information.
5. Summaries and AI answers as starting points that must lead back to identifiable sources.

Authority depends on the claim. A government registry may be decisive for a registration date and tell you little about daily operations. A company’s own website may be the best evidence of what it advertised on a particular date without establishing whether every advertised claim was true.

When Ten Sources Are Really Just One

Ten search results can give a claim an impressive appearance of credibility. But open the pages and you may find the same paragraph repeated across news sites and directories. If everybody copied one announcement, the claim hasn’t gained ten independent witnesses.

Look for where the story began. Publication dates can help, as can phrases repeated word for word. A later report counts as useful independent evidence when it adds something its own author checked, such as an original document or an eyewitness interview. A rewritten press release generally doesn’t.

AI search can make the repetition harder to spot. A system may retrieve several near-identical accounts and produce a confident summary that appears to cite independent reporting. Count the original investigations behind the sources before treating their agreement as corroboration.

Sometimes the best next search is one designed to challenge the story. Look for an original record that could contradict it. A good researcher wants to know where the confident explanation might break.

Build Search Queries Around the Evidence

Public research usually takes several searches because organizations change names, products and websites. Start with the identifiers behind the question. A trading name may appear on customer pages while a legal name appears in filings. A distinctive sentence can help you identify the earliest source of a copied announcement.

Google’s documentation covers operators such as site: and filetype:. You can use them to narrow a search.

• site:example.com "product name" looks for that wording on one domain.
• "company name" filetype:pdf can surface documents published as PDFs.
• "distinctive phrase" helps trace repeated or copied language.
• "company name" renamed or "company name" former address can lead to records of changes.
• site:gov.example "entity name" focuses on a specified public-authority domain when that domain is relevant.

Try historical names, alternate spellings and transliterations when appropriate. Different databases and search engines can expose different records, and their results change over time. An empty result tells you only that a particular query returned nothing. Google also warns that site: results are limited by indexing and retrieval, so they can’t establish that a page or fact doesn’t exist.

The Internet Remembers in Fragments

A website can change overnight while old copies linger elsewhere. Web archives help investigators compare versions of a page and learn what a site publicly displayed at particular moments. They’re especially useful when a current page no longer contains the wording you’re trying to check.

But a saved page doesn’t reveal the whole past. An archive might have captured a website months after a change, and it could have missed earlier versions entirely. The date the crawler saved a page is different from the date an event took place. Even a publication date visible on the page may have survived later edits.

The Common Crawl index provides another source of historical web records. Its collections are selective too. Absence from an archive is weak evidence that something never existed.

Some pages leave a rich history of captures. Others barely appear at all. Record what you found and the gaps you couldn’t close.

Make Sure the Records Belong to the Same Company

Company names are surprisingly poor identifiers. A brand may belong to a legal entity with a different name. Two unrelated businesses can share a similar name and operate in different jurisdictions. Combining their records without checking can produce a very convincing company history that belongs to nobody.

Official registries and filings help here. Look for a registration number and jurisdiction, then compare past names and dated entries where they’re available. A registry record can establish a legal company status as of a particular filing. It won’t necessarily tell you where staff currently work or whether an advertised service was delivered. OpenCorporates can help locate structured records across jurisdictions, though the authoritative local register should be checked for consequential claims.

Addresses need care too. Several companies may use one registered office, and a business can move without every directory updating at once. Keep conflicting details visible until you can resolve them. If the legal identifier, location and dates don’t line up, resist the urge to draw a connection from the name alone.

Domains and Public Internet Infrastructure Add Another Evidence Layer

Public technical records can help investigate a website’s history and relationships with other online properties. Domain registrations, DNS records, certificates and redirects reveal different pieces of that history. None automatically identifies the legal owner of the business behind the site.

RDAP provides standardized access to domain registration information and replaced WHOIS as the required system for generic top-level domain registration data. Available fields vary with the registry and privacy rules. Missing registrant details are simply unavailable evidence, and researchers shouldn’t invent a reason for their absence.

Choose technical records based on the question. DNS or nameservers can show a service’s infrastructure. Certificate records may identify publicly issued certificates for a domain. Historical DNS and redirect information can help establish changes over time. Shared hosting or a common CDN is weak evidence of common ownership because unrelated organizations often use the same providers.

Historical DNS tools can show where a domain pointed in the past. SecurityTrails is one example. Certificate indexes and internet scanning records answer different questions, and any historical observation needs a date. A service detected months ago may no longer be online.

Collect only the information relevant to the question. A long export of hosts and addresses doesn’t become a finding until you can explain what those records establish.

Documents Often Preserve Details That Websites Lose

Old PDFs and manuals can preserve product names, addresses and organizational details long after a website is redesigned. Search by file type when you need the original wording, then open the document itself. Inspect the publisher, visible date, revision number and any references that identify the version.

File properties can provide useful clues about authorship or editing history, but they are easy to misunderstand. A creation date may reflect an export, and an author field may come from a shared template. Treat metadata as context that needs corroboration, especially when someone copied or reprocessed the document.

How Investigators Find the Location Hidden in a Video

Look again at the Libya video from the opening. The footage showed a street and a mosque, and Higgins compared those details with a satellite map. The road’s shape and the location of the mosque’s minaret helped him narrow down the camera’s position. It was detective work based on ordinary things visible in the background.

That process is called geolocation, finding where an image or video was recorded. It can be more revealing than the caption attached to a clip. A genuine recording might circulate under a misleading location or be reposted years after the event. Check the original upload when possible, then compare recognizable features against independent maps or photographs.

Bellingcat’s toolkit collects resources for maps, image searches and archived material. They’re starting points for finding evidence. The convincing part of an investigation is the visible match and its limitations, rather than the number of apps used.

There’s another catch. Matching a location doesn’t establish who filmed the video or prove every claim made about the event. Every question needs its own evidence.

When a Neat Diagram Makes Weak Evidence Look Strong

It’s satisfying to draw a map of connected people and websites. A line from one node to another looks certain once it’s on the screen. But what created that line? Two domains using the same hosting provider can share a network address without sharing an owner. Two names mentioned in a document might have no relationship beyond appearing on the same page.

A relationship diagram is useful when every link can be traced to its supporting record. Label uncertain connections honestly. The diagram should help you investigate claims, rather than quietly turn guesses into facts.

AI Can Make the Wrong Story Sound Convincing

AI is useful for searching a stack of documents or organizing complicated notes. It can compare accounts and help you notice details you’d otherwise spend hours finding. The danger is how effortlessly it can turn incomplete evidence into a smooth explanation.

Two people with similar names may become one fictional biography. A report may contain nine paragraphs supporting a conclusion and one paragraph that challenges it. A generated summary can leave out the challenge and still sound reasonable.

The remedy starts with a simple question. Which original passage supports this sentence? Save the relevant document and the page or section where the evidence appeared. When the answer is consequential, inspect the original source yourself and keep the conflicting material available for the reader.

What Automation Helps With, and What It Can Obscure

APIs can retrieve records from registries and public datasets in a consistent format. Model Context Protocol (MCP) lets compatible AI applications discover and call external tools, including data lookups. The official MCP Registry lists published servers. These connections can save hours of copying fields between websites.

The returned fields still need interpretation. If an address is absent from a registry response, the value may simply be unavailable. An AI tool can turn that empty field into a plausible explanation without having evidence for one. Keep the original response, its source, the date and the exact claim you think it supports.

Tool permissions also matter. Reading an official record is different from uploading a file or changing an external system. Give a research workflow access only to the operations approved for its task. A report should distinguish what the tool retrieved from what the researcher concluded.

The Source Your Search Never Found

Search results are selected, not a complete record of everything published. An AI retrieval system makes its own selection when it chooses the documents a model receives. If a contradictory record was never retrieved, a careful sounding answer may still be misleading.

Our retrieval guide explains how that selection works. For an investigation, the practical response is to vary the search terms and look for older or original records. A missing result isn’t proof that the record never existed.

Keep the investigation proportional to the claim. A simple question may need only a few records, provided each one has a clear role. Save the source and date, check that the people or organizations in the records match, and look for a plausible alternative explanation. When evidence is contradictory, include that disagreement in the finding.

What CMX’s Old Chat Pages Can Tell Us

Our own website presents a useful problem for this kind of research. CMX began as a public chat project and now describes its work in website development, design, marketing, automation and OSINT. Yet some old chat pages remain accessible. Search can show a visitor the older identity long after the business has changed direction.

The exact date CMX stopped presenting chat as its main offering remains a separate research question. Answering it would require dated records from the transition itself. The current site can show what CMX publicly said when we checked it on October 10, 2026.

The services page described CMX’s current offerings. What’s Changed at CMX explained the change of direction, and The CMX Story described its history. The older Gaming Chat page showed why search engines and old links could still associate the company with chat.

WordPress records May 17, 2026 as the repurposing article’s original publication date, with another edit on October 8. The CMX Story carries an October 2025 publication date and an October 2026 edit. If a researcher quotes today’s wording as proof of what CMX said in 2025, the date on the page won’t support that claim.

The Wayback Machine can sometimes supply an earlier copy. Its capture date tells us when that version was saved, although the page could have changed before or after the capture. We didn’t establish a dated capture that settled CMX’s exact transition date, so the research should leave that date open.

Finding, checked October 10, 2026: CMX’s published website described its current digital services and acknowledged its earlier public chat platform. Historical chat pages were still accessible. These are first-party statements and pages controlled by CMX. They establish the site’s public account at the time of inspection, but they don’t independently verify client work or pin down the precise day the transition happened.

A corporate timeline built from today’s pages can easily put an old date beside a much newer claim. Checking what changed between versions may take longer, but it prevents a tidy history from becoming an inaccurate one.

Where OSINT, Search, SEO, and Digital Presence Meet

OSINT and SEO sometimes encounter the same public records. A search engine may display an outdated company description, while a researcher needs to establish when the original description changed and which source remains reliable. These are different questions, but a poorly maintained public record can create trouble for both.

Businesses can help by keeping official service descriptions, contact details, dates and profile links accurate. Correct important outdated listings at their sources and make current pages easy to identify. Clear first-party information gives researchers something to inspect and helps visitors find the right business.

Our guide to maintaining a digital presence covers that ongoing work. How Search Results and AI Answers Describe You explains how to investigate and correct inaccurate descriptions. This article concentrates on the evidence check that should happen before someone publishes or requests a correction.

Know When to Stop Looking

Public information still involves people. Keep personal details out of a report when they don’t help answer the question. A password or access token exposed in a search result is no invitation to use it. Follow an appropriate disclosure process and don’t test systems without permission.

There is also a point where another hour of searching won’t fix the evidence you lack. If the earliest available copy of a page was captured months after the event, say so. If two organizations share a registered address but you can’t establish a business relationship, leave that connection unresolved.

Our OSINT Field Guide covers tools and methods in more depth. If you need help verifying a public claim or an organization’s records, CMX also offers OSINT research services.