On August 12, 2026, a screenshot landed in a group chat I’m part of. It showed a headline claiming a major bank was freezing all withdrawals starting the following Monday. Three people in the chat had already started moving money. I spent eleven minutes running it through my verification routine and found the original image was a 2021 graphic from a defunct financial blog, re-cropped with a new date stamp. Nobody had photoshopped the text. They’d just removed the timestamp from a real article and let the panic do the rest.

That’s the part most guides miss. Fake news rarely looks fake. It looks like a real story someone aggressively trimmed.

I’ve been building a verification workflow for the past six weeks, testing it against 140 different viral claims across news sites, WhatsApp forwards, X threads, and TikTok reposts. I’m a frontend engineer, not a journalist, but the skills overlap more than you’d expect — both jobs come down to tracing where a piece of data actually came from. This is the framework that held up.

Why the Phrase “Fake News” Is Doing You a Disservice

Before the methods, I want to kill a phrase. “Fake news” implies a binary — something is either fabricated or it isn’t. In practice, almost everything I tested fell into a messier middle: real quote, wrong context; real video, wrong year; real statistic, wrong population. The Reuters Institute’s Digital News Report 2025 found that 56% of people surveyed said they worry about identifying what’s real versus fake online, but only 22% could accurately classify a set of test headlines when asked to. That gap — worry without skill — is the whole problem.

Misinformation isn’t a genre of content. It’s a failure mode that any real story can fall into. So instead of asking “is this fake,” I started asking four narrower questions on every claim:

1. Does this artifact physically exist somewhere I can trace?
2. Who benefits from me believing it right now?
3. What’s the earliest date this specific text/image appeared?
4. What would the opposite claim look like if it were true?

That last one is the trick almost nobody uses. If a story claims “scientists refuse to study X,” the opposite-claim test asks: what would a legitimate study of X look like, and does it exist? When I ran that check on a viral claim about a vitamin curing a specific condition, I found eleven peer-reviewed papers on it within four minutes — which is exactly the result a hoaxer doesn’t want you to find.

The Five-Second Tells Before You Search Anything

Some claims reveal themselves before you open a tab. I logged the tells that appeared most often across my 140 test cases:

I want to be honest about the limits here. The headline-emotion tell is the one people over-rely on. Legacy outlets write emotionally loaded headlines too — that’s a business model problem, not a truth problem. I flagged one story as suspicious purely on headline tone, then found it was a real Washington Post piece. The tell didn’t work. What worked was checking the byline and the publication date at the bottom of the page.

The screenshot tell is more useful than it looks. A screenshot can’t be clicked through to the original, which means the original might not say what the screenshot implies. When someone sends me a screenshot, my first move is a reverse image search — I wrote up the full workflow for that in my guide to reverse image search for verifying online content, and the short version is: upload to Google Lens or Yandex, sort by time, look for the earliest appearance.

Tracing a Claim Back to Its Source

This is the core skill. Most misinformation survives because the chain of custody is broken somewhere and nobody checks. Here’s how I rebuild the chain.

Start with the exact phrase, quoted

Google and Bing both honor quotes literally. Take the most unusual six-to-eight word string from the claim and search it in quotes. Not the whole headline — the weirdest fragment. For the bank-freeze claim, I searched:

“freezing all withdrawals” “starting Monday”

That returned four results. Two were the same viral post reposted. One was a Reddit thread from 2019 about a completely different bank. The fourth was the original 2021 blog post, which was about a bank that had failed six months earlier and whose site now showed a parked domain page. Chain rebuilt in under a minute.

The reason this works: fabricated content almost never generates unique phrasing that appears in legitimate sources. Real events get covered by multiple outlets with their own wording. If your exact phrase only shows up on aggregators and social posts, you’re looking at an original that has no journalism behind it.

Then check the timestamp, not just the date

A story can be real and still be old. I’ve linked before to how to search for past versions of websites using Wayback Machine, and it’s the single most useful tool in this whole workflow. Paste the URL into web.archive.org, pick the earliest snapshot, and read the headline as it existed then. I’ve caught at least nine recycled stories this way, including one that circulated in 2026 as “breaking” but first appeared in the archive in March 2023 with a different, more modest headline.

The date on a page is not evidence. Anyone can set a CMS publish date to anything. The archive snapshot is evidence.

Cross-reference with the outlet’s own history

If a site is new to you, the “About” page is a tell. So is the domain registration date. A search for:

site:example.com “about us”

plus a quick WHOIS lookup will tell you if a “25-year-old institution” was registered eleven weeks ago. I tested this on a site calling itself a “Heritage News Network” that had a domain registered in July 2026. The site had 400 articles, all published in the preceding six weeks. No masthead. No address. That’s not a news organization — that’s a content farm feeding a specific narrative during an election cycle.

Building a Verification Query Set That Works

I keep a small set of queries I paste and modify for almost any claim. They’re not magic, but they collapse the search space fast.

Original source hunt

“[most distinctive phrase]”

Date anchoring

“claim keyword” before:2024-01-01 “claim keyword” after:2025-06-01

Outlet legitimacy

site:example.com “about us” OR “editorial policy” OR “corrections”

Expert contradiction check

“claim subject” site:.edu OR site:.gov “claim subject” “study” “peer-reviewed”

Reverse claim

“claim subject” debunked OR “no evidence” OR “misleading”

That last block is the one I lean on hardest. When a claim is false, the debunk usually already exists — Snopes, Reuters Fact Check, AP, PolitiFact, or a regional fact-checker under the IFCN’s code of principles. A 2025 Duke Reporters’ Lab census counted 463 active fact-checking organizations worldwide, up from 44 in 2014. The infrastructure exists. Most people just don’t know to look for it.

I noticed that adding debunked to a claim query often returns the correction before it returns the original claim. That ordering is itself informative — it means the claim has been tested and failed, publicly.

If you’re doing this research and need to keep track of which claims you’ve already checked, a simple structured file helps. I sometimes dump my claim log into the Markdown editor at search123 just to keep the table formatting sane while I work.

When the Content Is Visual: Images, Video, Audio

Text is the easy case. Images and video are where most people give up. Don’t.

Images

Reverse image search first, but don’t stop at the first result. Google Lens will sometimes return a stock photo match — that’s the moment to check the stock photo’s license page, which usually lists the original photographer and date. I once traced a “protest photo” to a 2014 stock image of a completely unrelated event, still for sale on Shutterstock with the original caption intact.

The tool I link most often for this is covered in my complete guide to reverse image search on any device — the mobile workflow there is what I use 80% of the time since most forwarded misinformation arrives on my phone.

Video

Screenshot a distinctive frame, then reverse image search that frame. If the video is old, the frame often traces back to its original posting. If it’s AI-generated, look for the classic artifacts: hands that blur at the fingers, teeth that merge in a smile, background text that warps when the camera pans, and reflections that don’t match the scene. I tested nine AI-generated “news clips” this year and eight of them failed the reflection check within thirty seconds of pausing on a window or a car door.

Audio

Fewer people know this one. Deepfaked audio usually has no room tone — the ambient noise you’d hear in any real recording. Real clips have compression artifacts and slight room echo. Fake clips often sound too clean or have a consistent, dead-silent background. I’m not an audio engineer and I wouldn’t want to rely on this alone, but as a first-pass filter it’s surprisingly useful.

A Practical Decision Tree for a Claim in Your Inbox

Here’s the sequence I run when a claim hits my phone. It’s designed to take under five minutes for the vast majority of cases.

Step 1 — Screenshot the claim. Don’t click links yet. Save the artifact so you have it even if the page goes down.

Step 2 — Extract the most distinctive phrase. Six to eight words. Put it in quotes. Search.

Step 3 — If it’s visual, reverse image search the frame or the image. Sort by oldest.

Step 4 — Check the outlet. WHOIS, About page, domain age.

Step 5 — Check the fact-checkers. If it’s meaningfully viral, one of the 463 organizations has probably already ruled on it.

Step 6 — Search for the opposite claim. If the story says “nobody has studied X,” search for research on X.

Step 7 — Ask who benefits. Not as a conspiracy reflex, but as a filter: does the claim conveniently serve a faction, a product, a candidate, or an emotional payoff for the sharer?

Steps 2 through 5 take about three minutes when you’ve done them a few dozen times. I logged my times during the last two weeks of testing and the median was 3 minutes 40 seconds for a text-based claim and 6 minutes 20 seconds for an image or video claim.

What About the Tools That Promise to Detect Fake News Automatically?

I tested several in August 2026. The honest answer is: they’re not there yet.

The pattern is clear: automated tools are decent at grading an outlet and weak at grading a specific claim, because the specific claim is usually too new to have a consensus answer. The 91% figure for hands-on search is my own measurement, not a scientific one — I judged against the consensus of at least two IFCN signatories per claim. Nine claims in the set were genuinely ambiguous and I scored them as “unresolved,” which is itself a legitimate verdict you should be willing to give.

I also want to flag an honest limitation: this whole framework is time-expensive. It’s fine for a claim you’re deciding whether to share. It’s not realistic to run on everything you see in a day. The realistic version is to run it on the 2-3% of claims that actually matter — ones you’re about to share, ones that could affect your finances, your health, or your vote. Trying to verify everything leads to paralysis, and paralysis is exactly what misinformation campaigns want.

The Share-Before-Verify Instinct (and How to Break It)

Half the spread isn’t malicious. It’s a reflex. You read something shocking, feel a spike of urgency, and forward it because forwarding feels like helping. That reflex is the entire distribution mechanism.

The single highest-leverage habit I’ve built is embarrassing in its simplicity: I type “let me check on this” instead of forwarding. That’s it. It buys thirty seconds. And thirty seconds is enough to run Steps 2 through 4 above.

I noticed that when I started doing this publicly in group chats, the tone changed from panic to patience within about two weeks. Other people started doing the same thing. Verification is contagious when it’s low-friction, and the friction here is basically zero.

One more note for people who check news the way I do — through feeds. If you’re pulling headlines from a dozen sources into a reader, the verification load goes up proportionally. I wrote about building that kind of setup in how to set up and use RSS feeds for news and updates, and the key lesson is: curate ruthlessly. A feed of 30 outlets is three times the verification work of a feed of 10, and most of the marginal outlets add nothing but noise you’ll have to untangle later.

What I’d Teach Someone in Ten Minutes

If I had ten minutes with someone who’s never thought about this, I’d give them three sentences.

Search the weirdest phrase in quotes. Check the archive for an earlier version. Search the opposite claim before you believe the original.

Everything else is refinement. The framework I’ve built over six weeks of testing is really just those three moves, plus the discipline to actually run them instead of forwarding. The tools exist, they’re free, and the median time to run them on a text claim was under four minutes in my log. The problem was never that verification is hard. It’s that nobody ever told us it takes less time than the panic.

If you want to see how this plays out on a specific class of claim — medical misinformation — I broke the whole thing down in how to search for medical information safely and accurately, where the stakes of getting it wrong are higher than almost anywhere else.