I saved over fifty short videos full of SEO advice. Here is what each one claimed, what the creator was selling, and what held up when I checked. The claims that can be tested are being tested on real sites, and the results land here, dated, whichever way they go.
False8
Overstated4
Testing4
Partly true3
Unproven3
Holds up1
Won’t test1
The loudest claims are the ones that fall apart. The useful advice is quieter, and
most of it agrees with what Google has already written down.
What the verdicts mean
Holds up
Checked against the source, and it is right.
Partly true
The core is right. A detail or the size of the effect is not.
Overstated
Something real, sold as bigger than it is.
False
The source says otherwise.
Unproven
Cannot be checked from the outside yet.
Testing
Being tested on a real site now. See the tests below.
Won’t test
Against Google’s rules or the law, so not worth risking anyone’s domain.
These grade whether a claim is true. The clip library
grades a different question, whether a clip is worth your minutes, so its three
verdicts are not these seven.
How I check, and how I credit people
I save short SEO and AI-search videos from TikTok and Instagram. Each one gets
transcribed, and every claim in it is pulled out and checked against the original
source: Google’s own documentation, the study being quoted, or the tool itself.
A claim that can be tested on a real site gets tested, with a page left alone for
comparison, success criteria written down before the data exists, and a dated log.
Grading a claim is not grading a person. Most of these creators post useful things
alongside the hype, and every row links to who said it. Where I saved the clip by
link, the name goes to that exact post, so you can watch it and disagree with me.
Where I saved it off my phone there is no link to the post, so the name goes to the
creator’s profile instead and is marked (profile). A few clips lost
their handle entirely; those stay unlinked rather than guessed at.
If I got something wrong, tell me and I’ll fix it and log the change.
Push new pages to Bing with IndexNow and they get picked up faster, which matters because ChatGPT leans on Bing.
The setup
Two new Kamloops sites launch the same week. kamloopsdailynews.com pings IndexNow whenever new headlines land, at most every six hours. kamloopsjustlisted.com launches without IndexNow for its first four weeks, then gets it.
What gets measured
The date each home page first shows up in Bing and in Google (Bing Webmaster Tools and Search Console, checked daily), and how many pages are indexed each week.
What counts as a win
IndexNow counts as working here if the site that pings it is in Bing at least a week sooner. If both land within a few days of each other, the claim is overstated for a small site.
What could fool us
These are different sites with different content. The directory domain came with a few real links from its previous owner, and the news domain came with only spam links. Both are logged so the result can be read with that in mind.
Running since 19 September 2026
Do three-question FAQs get a page cited by AI?
The claim
Add three questions and short answers to the bottom of your top pages and AI tools start citing them.
The setup
Three wiredbuilder.com pages get a three-question FAQ, written the way someone would ask an AI. Three similar pages stay exactly as they are. All six are picked from Search Console traffic before anything changes.
What gets measured
Search Console impressions and clicks for each page, plus the same set of questions asked every week in ChatGPT, Perplexity and Google AI Mode, recording which pages get cited.
What counts as a win
Over eight weeks, one of the three FAQ pages gets cited in an AI answer where none of the three untouched pages do. Impressions get recorded alongside, as context, and are not the verdict. That is a narrower bar than I first wrote, and the reason is below.
What could fool us
This site takes about four clicks a quarter from search, so an impressions comparison here would measure noise and nothing else. I checked whether a sister site was better and living-wired.com is quieter still. A Google update during the test would hit all six pages at once. Any other edit to any of the six voids that pair, and gets logged here.
Choosing the test site
Does an llms.txt file change what ChatGPT recommends?
The claim
Add an llms.txt file and ChatGPT starts recommending the business.
The setup
One established site that has no llms.txt yet. Two weeks of baseline first: twenty buyer-style questions asked in ChatGPT, Perplexity and Google AI Mode, with every answer saved. Then add the file and change nothing else, and ask the same questions every week for six weeks.
What gets measured
How many of the twenty questions mention or cite the site, per tool, before and after.
What counts as a win
The site shows up in more answers after the file than before, in ChatGPT or Perplexity. Google says the file does nothing for its AI features, so Google AI Mode should not move. If it does, that is news too.
What could fool us
AI answers change from one run to the next, so each question is asked three times. Model updates during the test get logged.
Running since 12 August 2026. First readings 19 September
Does a Google Business Profile Offer bring in paid audits?
The claim
Post an Offer on your Google Business Profile with a coupon code, and local searchers find it and redeem it for free.
The setup
My website audit sells for $50. A Google Business Profile Offer post carries a coupon that drops it to $10, and the redeem link points at a dedicated page with its own tracking parameters so the source is unambiguous. The $10 is a filter, not a revenue line: Stripe keeps about 59 cents of it. Nothing else about the audit changes while this runs.
What gets measured
Views on the Offer post from the profile’s own insights, sessions on the redeem page from that link, coupon redemptions at checkout, and how many of those $10 buyers become a real project after my follow-up.
What counts as a win
Any redemption traceable to the Offer post beats the alternative, which is that the post is decoration. The bar I actually care about is one project out of the first ten redemptions, because that is what makes the funnel worth keeping.
What could fool us
One business profile in one city, so this is a single data point and not a rule. The coupon can be found and shared outside the Offer post, which is why the tracking parameters matter more than the coupon count. A quiet month locally looks the same as a mechanism that does not work.
Every claim so far
24 claims so far. Search them, or filter down to one verdict.
The clips that open with “SEO is dead” tend to close with “comment a word for my guide”. Three in the pile say the words outright; the rest argue a softer version.
checked
False
Google’s own guide says its AI features are “rooted in our core Search ranking and quality systems.” Neil Patel’s figures across 22 companies put AI search at 7.4% of leads, up from 3.1% a year earlier. Growing fast, not replacing search.
Both are reading Google’s guide. Williams-Cook’s clip was the only one in the pile that needed no correction. @theseotips founded Arvow, which sells an AI content tool.
checked
Holds up
Google, in writing: optimizing for generative AI search “is optimizing for the search experience, and thus still SEO.”
Checked 7 Sep 2026
claimed
You need llms.txt, Markdown copies, chunked content or special schema to show up in AI answers.
Yalif’s four-part frame is useful. His schema statistics come with no source.
checked
Falsefor Google
Google: “You don’t need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search,” and structured data “isn’t required.” Other AI tools read the web differently, which is why llms.txt is one of the tests on this page.
Checked 7 Sep 2026
claimed
Adding an llms.txt file got a vehicle-wrap shop recommended by ChatGPT.
IndexNow does push new pages to Bing, and Bing Webmaster Tools imports a site straight from Search Console. ChatGPT leaned on Bing’s index for years, but OpenAI now runs its own crawler too. How much faster IndexNow really is gets tested below.
Wrapped around the sensible IndexNow advice above.
checked
Falseand closer than it was
When I first checked this the best figure was SparkToro and Datos for 2024: about 58.5% in the US and 59.7% in the EU. SparkToro’s 2026 numbers put it at roughly 68% of Google searches ending without a click. Still not over 80, so the claim stays false, but it was wrong by twenty points and is now wrong by twelve. Worth saying plainly, because the direction of travel favours the person who made the claim.
Checked 20 Sep 2026
claimed
“Google confirmed” AI answers now appear on over 40% of searches.
Google never said it. The figure is vendor panel data (Similarweb 40% and up, BrightEdge about 48% in February 2026), and estimates run from the low 20s to the high 40s depending on which searches are measured.
Google says its AI features are built on its ranking systems, so ranking still decides what gets pulled in. The shift from single keywords to clusters of questions is real.
Checked 7 Sep 2026
claimed
Claude Code just made most SEO agencies irrelevant.
Hidden text is named in Google’s spam policies. It is an old trick that can get a site penalized.
Checked 7 Sep 2026
claimed
Put “reddit” in your URL, title and H1, and you’ll get pulled into AI answers.
A short video crediting SEO David Quaid
A tactic video, no product attached. I saved it off my phone rather than by link, so I have the 46 seconds and no idea who posted it. Named here as the clip credits it, and nothing more.
Told as SEO history, and the history is mostly accurate.
checked
Overstated
The 2023 DOJ testimony and the 2024 leak show Google collects click data and uses it in systems like Navboost. That is not the same as “get people to click your result and you’ll rank.”
Checked 7 Sep 2026
claimed
Buy an expired domain, rebuild it from the Wayback Machine, and link from it to your site.
Google’s spam policies have named expired domain abuse since March 2024, and republishing someone else’s archived site is a copyright problem. Not something to run on any client’s domain.
Checked 14 Sep 2026
claimed
ChatGPT cites pages that match the searches it runs, and you can read those searches in your browser.
The feature is real: Google’s agentic calling (“Have AI get prices”) started in Search Labs in January 2025, and Google announced it for AI Mode in the US at I/O 2026. The second half is the part that isn’t established. That the business which answered got recommended over better-ranked ones rests on one search filmed on a phone, with the site never named. The mechanism makes it plausible. One clip does not make it measured.
Checked 19 Sep 2026
claimed
A site published 50,000 AI pages and only about 130 got indexed.
Read off a site: search, and the site isn’t named. It matches Google’s guidance against mass-produced pages, which is why pages here get published in small batches with the indexed count watched after each one.
The paper (Aggarwal et al., arXiv 2311.09735) does say “up to 40%”, and that results vary by topic. Adding citations, quotes and statistics helped; keyword stuffing did not. The clip’s 132% figure isn’t in the paper’s abstract.
Checked 7 Sep 2026
claimed
AI tools cite the same few pages for each topic, so get your brand onto those pages.
Uses a paid citation tracker (peec.ai) to find the pages.
checked
Unproven
Plausible, and testable one niche at a time. The clean version is asking for honest mentions. Paying for undisclosed listicle spots or dropping Reddit comments is not.
Checked 7 Sep 2026
claimed
Post a Google Business Profile Offer with a coupon code and local searchers find it and redeem it.
The mechanism is real and checkable in the product: a Business Profile post can be typed as an Offer rather than an Update, and the Offer type carries a coupon code field and a link to redeem it. Those two fields map onto Stripe promotion codes with no custom code, which is why this is cheap to try. Whether anyone actually finds the post and buys is the part nobody has shown, and the part I am testing.
Checked 19 Sep 2026
The log
The log
Date
What changed
Started the Reddit interception experiment, the first claim from a saved video tested on a live site. It has its own page.
Went through 61 saved clips in one evening. Each was transcribed, and every claim in it was checked against its source: Google’s documentation, the study being quoted, or the tool itself. Most of the rows above come from that pass.
Added the expired-domain tactic and the “Google confirmed 40%” claim.
Started this page. Picked three new tests to run alongside the Reddit one: IndexNow, three-question FAQs and llms.txt.
Went back through every row against the saved clip and the notes I wrote when I checked it. Swapped each creator link from their profile to the exact post, for the clips I saved by link. Twenty-three of the twenty-eight now point at the post itself. Three I had screen-recorded off my phone instead, so there is no post URL to point at: those keep a profile link, marked as such. Two lost their handle in the recording and stay unlinked rather than guessed at. No platform was wrong. Edward Sturm’s clips are on Instagram and the rest are on TikTok, which is what the rows already said.
One verdict changed, and it is mine that was wrong, not his. I had Matt Diamante’s agentic-calling clip down as “Holds up”. Google’s feature is real and I had checked that properly. The part I waved through is the half that actually sells the clip: that the shop which answered the phone beat better-ranked shops. That rests on a single search filmed on a phone, with the business never named. Feature confirmed, ranking claim not, so the row is “Partly true” now. My own note from 7 September already graded it that way and I did not read my own note carefully enough.
evening
Both pages merged and live. That started the FAQ test, so I took its day-zero AI reading five minutes after the deploy landed, while no engine had recrawled: six questions, three runs each, on a grounded AI surface. Not one answer cited any wiredbuilder.com page, treatment or comparison. Even asking who I am by name returns nothing. A floor of zero is a clean place to measure from, if a slightly bleak one.
narrowing the FAQ test
Changed the success criterion for the FAQ test before any data existed, which is the only honest time to change one. It used to say the FAQ pages should gain more impressions than the untouched ones, or get cited where they do not. The impressions half cannot work here: this site takes about four clicks a quarter from search, and the busiest page in the test sees eighty impressions in ninety days. Any eight-week move in that is weather, not signal. I checked whether living-wired.com was a better host and it is quieter still. So the criterion is now citations only, and impressions are recorded as context. The test got narrower and more honest at the same time.
first readings on the Reddit experiment
Five weeks late, and they are interesting. The tactic page ranks 4th on its reddit query, so the first success criterion is met. The control page, which has no reddit anywhere in it, ranks 8th on that same query and 1st on its own. A grounded AI retrieves the tactic page for the Reddit question and the control page for the pros-and-cons question, so both pages are inside the machine, and neither gets named out loud. The tactic works and has not yet been shown to beat an ordinary good page. Full numbers and the mistake I made scoring them are on the experiment page.
a fifth test
Added the Google Business Profile Offer to the list, because I am now running it rather than just reading about it. My website audit is $50, an Offer post on the business profile carries a coupon that makes it $10, and the redeem link is tagged so I can tell what came from the post. The claim that a free post on a profile brings in paying local customers is exactly the sort of thing this page exists to settle.
Checked the last nine claims sitting in the clip library’s backlog, and two of them changed what I think about a test on this page. The first: a creator promising that publishing posts as markdown brings hundreds of AI crawler hits within hours. Measured studies find close to the opposite, including one test logging zero visits to .md files against 137 to the HTML, and an analysis across 300,000 domains finding no link between these files and AI citations. That is the same family of claim as the llms.txt test below, and it means that test now has a likely answer before it starts. I am running it anyway, because a likely answer is not a measured one and my own site is cheap to experiment on, but the expectation has moved.
Sources: OtterlyAI, Markdown vs HTML test, SE Ranking, llms.txt across 300,000 domains
a number moving against me
Updated the zero-click row rather than leaving it flattering. I graded "over 80% of searches end with zero clicks" false on the strength of SparkToro’s 2024 figure of about 58.5%. Their 2026 figure is about 68%. The claim is still false, and it is a lot less false than it was. Leaving the old evidence in place would have made my correction look stronger than it is.
Sources: SparkToro 2024 study, SparkToro 2026 study
Where it stands
The pattern so far is plain. The loudest claims (“SEO is dead”, “80%
of searches are zero-click”, “Google confirmed”) fall apart when you
check the source. The useful advice is quieter, and most of it agrees with what Google
has already written down. The tests above will settle the rest.
Get the next result.
I email when a test on this page posts a verdict, whichever way it lands. Roughly twice a month, and nothing else.
One email at a time, unsubscribe whenever. See the Privacy Policy.
Seen an SEO claim you want tested?
Send it here. I check it against the source first. If it is worth
testing, it goes on the tracker with a dated verdict, and you get the
result by email.
Have your own site checked this way
The same habit, pointed at one site: what is actually wrong, what the
evidence says, and what I would do about it in what order.
