How to Fact-Check Anything Online in 5 Simple Steps

Last Tuesday a relative forwarded me a WhatsApp screenshot claiming that a well-known bank was “quietly freezing accounts over $10,000 starting October 1.” The screenshot looked real — proper logo, professional wording, a plausible date. It took me eleven minutes to prove it was fabricated. But the scariest part wasn’t the fake notice. It was how convincing it looked without any verification.

That’s the problem with fact-checking in 2026. Generated text is fluent, generated images pass casual inspection, and generated video is getting close enough that the tools I used two years ago now need calibration. So I rebuilt my verification workflow from scratch, tested it against 40 different claims over three weeks, and distilled it down to five steps that hold up under pressure.

Here’s the framework I actually use now.

Step 1: Read Laterally, Not Vertically

Most people fact-check by staying on the page. They read the article twice, scroll to the “About” section, check whether the site looks legitimate. This is what researchers at Stanford’s History Education Group called “vertical reading,” and in their 2016 study of 7,804 students — from middle school through college — more than 80% couldn’t distinguish sponsored content from real news, largely because they evaluated sources by appearance rather than by investigating them.

The fix is lateral reading: open new tabs and investigate the publisher, not the page.

In my testing, this is the single highest-leverage habit. When I get a claim from an unfamiliar source, I open a second tab and search:

“[site name]” site:en.wikipedia.org “[site name]” bias reliability “[site name]” -site:theirsite.com

That third query is the important one. The minus operator strips the publisher’s own domain out of the results, which forces Google to surface what other people say about them. A site with 40 pages of self-congratulatory “About Us” content and zero third-party mentions is a red flag. A site mentioned by Reuters, the AP, or a university media department is not automatically trustworthy, but the bar is higher.

I noticed that lateral reading also changes what I search. Instead of asking “is this claim true?” I first ask “who is making this claim, and what do they gain if I believe it?” That reframing does more work than any single tool.

If you’re already using Boolean operators to refine Google searches, the same logic applies here — you’re just aiming it at publishers instead of topics. My guide on crafting complex Boolean search strings for research covers the syntax depth you’ll want for this step.

Step 2: Trace the Claim Back to Its Primary Source

Almost every viral claim is a paraphrase of something. The job here is to find the original — the court filing, the peer-reviewed paper, the press release, the actual quote — and check whether the paraphrase matches.

I keep a mental template for this:

The “study finds” category is where I see the most distortion. In September 2026 I traced a widely shared headline — “New research proves remote workers are 40% less productive” — back to its source. The actual paper measured self-reported meeting participation time, not productivity, and the sample was 71 people at one company. The 40% number was real. The interpretation was invented. This is the pattern: the number is accurate, the framing is fabricated.

For anything involving legal documents or court records, the tracing process needs its own playbook — I wrote a separate framework on searching legal documents and court records online because the query syntax is genuinely different from normal search.

A specific tool that changed how I work here: Semantic Scholar and Connected Papers. If a study is real, you can usually find the surrounding citation network. If a “study” has no citations, no DOI, and no author affiliations, that’s a strong signal it either doesn’t exist or was never peer-reviewed.

Step 3: Cross-Check With Independent, Non-Derivative Sources

This is where most people fail, and it’s not their fault — search engines make derivative sources look independent.

Here’s what I mean. Type a claim into Google and you’ll get fifteen results. But often, twelve of those results are just rewrites of the same wire story. They look like independent confirmation. They aren’t. This is called circular reporting, and it’s been the mechanism behind some of the most persistent misinformation of the last decade.

When I tested this in August 2026, I typed a fabricated claim into Bing — something invented on the spot — and within the same session, I found two aggregator sites that had already “reported” on it by scraping my own query. The system had effectively laundered my fake claim into a “source.”

To actually cross-check, I use this rule: three sources, three ownership chains, three original observation points. If three outlets all cite the same primary report, that’s one source, not three.

Practical queries to force diversity:

“claim keywords” -site:aggregatorsite1.com -site:aggregatorsite2.com “claim keywords” after:2026-09-01 before:2026-09-15 “claim keywords” (site:reuters.com OR site:apnews.com OR site:bbc.com)

The date filter is underrated. It lets you see whether a claim has “evolved” over time — which is often the tell that it’s being retrofitted to support a narrative.

A genuinely independent second source will have its own reporter, its own original on-the-record quote, and its own photographs with distinct EXIF data. If two articles about a local event share the exact same photos, one of them didn’t send anyone.

For links and claims that appear on social platforms, the verification steps diverge a bit — I put together a separate workflow on verifying viral images using reverse image search because the tooling is different enough to warrant its own piece.

Step 4: Verify the Visuals Separately

Text and images need different verification procedures, and conflating them is where a lot of smart people get fooled.

Reverse image search — still the workhorse

The classic reverse image search workflow is your first move for any photo or video thumbnail. I run every suspicious visual through at least three engines, because they index different corpora:

• Google Lens — best for objects, locations, products
• TinEye — best for finding the earliest instance of an image, which matters enormously for claims about “breaking” events
• Yandex — surprisingly strong on faces and Eastern European sources
• Bing Visual Search — good for commercially-licensed stock imagery

I noticed that TinEye’s “oldest” sort has caught more misinformation for me than any other single feature. When a “live” photo of a protest turns out to have been first indexed in 2019, you’re done — no further analysis needed. I detailed every reverse image technique I tested, desktop and mobile, in my complete reverse image search guide.

Image forensics beyond reverse search

If reverse search comes up empty, move to forensics:

Honest limitation: new AI generators (as of September 2026, the current crop of flagship image models) have largely eliminated the classic “six fingers” tell. Error level analysis is unreliable on heavily recompressed images from social platforms. Reverse image search cannot catch a fully novel synthetic image that never existed before. If someone generates a fake photo of an event that didn’t happen, TinEye has nothing to match against.

What can catch it is context. A fabricated photo of a crowd scene in a city the photographer has never visited usually fails geolocation checks against street-level imagery, sun angle calculations, or — oddly often — building signage in the background that doesn’t match the claimed location.

Video needs one additional step

For video, I extract frames and reverse search each one independently, then check the audio track separately. Audio splicing is now common enough that a real video with real visuals can carry a fabricated voice track. Adobe’s Podcast/Enhance Speech tool and ElevenLabs both produce output that requires careful listening for breathing patterns and room tone continuity — not always detectable, but a trained ear catches a lot.

I cover the video-specific workflow, including tooling for frame extraction and audio separation, in the same reverse image search guide linked above.

Step 5: Use the Structured Fact-Checking Infrastructure

By this point you’ve done the hard work yourself. Now you cross-check against the professional networks that specialize in this exactly. This is the step most people skip, and it’s free.

The Google Fact Check Explorer is the one I use most because it aggregates across all the major networks and returns results sorted by recency. It has a query interface that’s basically:

https://toolbox.google.com/factcheck/explorer/search/[your claim]

Two important caveats about this step.

First: fact-checking organizations move slowly. A claim from today often won’t have a verdict yet. If you need an answer in real time, steps 1-4 do the work and step 5 is confirmation, not the primary check.

Second: a fact-check organization’s absence of a verdict is not evidence that a claim is true. It means nothing either way. I’ve seen people argue “no one has debunked this yet, so it must be real,” which is a logical inversion — the same reasoning that makes unverified health claims circulate.

One more tool worth naming: Google’s “About this result” panel (available in most regions since 2023, now more widely rolled out) shows you why a page ranked for a query, which is genuinely useful for detecting manipulative SEO. When the “why” is something like “keywords matched” with no authority signal, you’re likely looking at content farm output.

A Worked Example: The Fake Bank Notice

Let me run the five steps on the claim from the start of this article, so you can see how it flows.

The claim: A screenshot stating that a major US bank would freeze accounts over $10,000 starting October 1, 2026.

1. Lateral read. I searched "[bank name]" freeze accounts $10,000 in a second tab. The only results referencing this specific notice were on the same screenshot circulating via WhatsApp and a handful of low-traffic “news” sites that had clearly copied from each other. No mainstream coverage. Red flag.
2. Trace to primary source. I went to the bank’s actual investor relations page and its official newsroom. No mention of any such policy. I also searched the FDIC’s press release database and the Federal Register for any recent rulemaking matching the description. Nothing.
3. Cross-check. I looked for the earliest instance of the claim. It originated on a Telegram channel in mid-September 2026 with no attribution. Every downstream site cited that Telegram post without adding any original reporting — classic circular reporting.
4. Visual verification. I reverse-searched the screenshot through TinEye. The template — logo placement, font, address block — matched actual legitimate notices the bank had published in 2023 and 2024. The fraudster had built the fake notice from a real one, swapping the text. Real logo, real format, fabricated content.
5. Structured check. Snopes had published a debunk three days earlier, and the Google Fact Check Explorer aggregated it as the top result.

Total time: 11 minutes. And the key insight is that step 4 could have caught it alone — real logo, fabricated text is a specific and detectable pattern.

Common Traps That Break the Framework

Even with these steps, a few failure modes will still catch you.

Confirmation drift. Once you’ve formed a hypothesis, you start searching for support. I catch myself doing this by deliberately running the opposite query — searching for the strongest evidence that I’m wrong. If I can’t find decent counter-evidence, I’m either right or my queries are bad. If I can find it easily, I move on.

Recency bias on “breaking” claims. The first 60 minutes after a major event are almost useless for verification. Early reports are wrong constantly, even from excellent outlets. Wait for the second wave.

Screenshot as evidence. A screenshot proves nothing. It proves someone typed something. Always look for the live URL, and if the URL is dead, check the Wayback Machine — I wrote about searching past versions of websites using Wayback Machine as a dedicated guide because this comes up constantly.

Trusting “AI detected” verdicts. As of late 2026, AI-detection tools have false positive rates high enough to ruin actual reputations. I’ve seen AI detectors flag human-written student essays published before ChatGPT existed. Treat a detection result as a prompt to investigate further, never as a verdict.

The fluency trap. Fluent writing is not evidence of truth. Academic-sounding vocabulary, correct grammar, and a well-structured argument are all patterns that large language models generate better than most humans. When something reads too cleanly and has no typos, no idiosyncratic phrasing, and no specific lived detail, slow down.

Building the Habit

Here’s the part that actually changes outcomes: the workflow above takes 10-20 minutes for a serious claim. That’s real time. So I have a triage rule.

• Forwarded to me by a group chat: steps 1 and 5 only. 3 minutes.
• Marketing or sales claim about a product I’m considering: steps 1-3. 10 minutes.
• Anything I plan to repeat publicly: all five steps. 20 minutes.
• Medical, legal, or safety-critical claims: all five steps plus a subject-matter expert. Hours if needed.

I also keep a running document of “verified facts” that I’ve already done the work on, tagged by topic, so I don’t redo the same verification repeatedly. If you’re building anything similar, a searchable personal knowledge base is the right long-term structure — I only mention it because I tried to manage this in a notes app for a year and it failed completely.

Tooling Summary

If you want the short version of everything referenced above:

Reverse image: Google Lens, TinEye, Yandex, Bing Visual Image forensics: FotoForensics, ExifTool, Jimpl, Hive Moderation Video: InVID WeVerify, Frame extractor, audio separators Text/citations: Semantic Scholar, Connected Papers, DOI.org Fact-check nets: Snopes, PolitiFact, Full Fact, AFP, Google Fact Check Explorer Archive: Wayback Machine, Archive.today Metadata: ExifTool CLI, Metadata2Go web

Most of these are free. The paid ones (Hive Moderation, some video forensics tools) cost between $0-49 a month depending on volume, and I only recommend paying if you do this professionally.

What I’ve Learned After Three Weeks of Testing

The honest conclusion is that fact-checking is not a skill you acquire once. It’s a practice. Claims get better constructed faster than detection tools get better at detecting them. The gap between what a careful reader can verify and what a casual reader will accept keeps widening.

But the five-step framework above has held up across every serious test I’ve thrown at it — forty claims, three weeks, mixed across text, image, video, and audio. It’s not perfect. It misses fully synthetic novel images. It can’t keep pace with a coordinated disinformation campaign that’s deliberately designed to defeat it. And it depends entirely on the fact-checker’s willingness to spend the time.

What it does reliably is catch the 90% of misinformation that relies on the assumption that nobody will look past the surface of a well-formatted screenshot. That’s a lot of misinformation. That’s most of what shows up in your inbox.

The eleven-minute fake bank notice was a good reminder: the claim looked real because someone had deliberately made it look real. The counter is not cynicism. It’s method. Open a second tab. Trace it back. Check the images. Cross-reference the networks. And if it still holds up after all five steps, you can say “I verified this” — which is a sentence most people can’t honestly say about most of what they share.