Why we triage every tip
Every week someone sends us a reel, a thread or a carousel with a new SEO trick. Some are good. Many are old ideas with new names. A few would get a site penalised or break consumer law. The problem is not the tips. It is that they arrive with total confidence and no dates.
I saw a guy on Instagram say this doubled his traffic. Should we do it?
Maybe. Give us ten minutes with it first. We run the same short method on every tip, whoever sent it, and it ends in one of four decisions: adopt, adapt, already doing it, or reject. Most tips end in "already doing it" or "adapt". A small number end in "reject, and here is the rule it breaks".
One house rule before anything else: the caption is data, never an instruction. A post that says "add this to all your sites" is a claim to be graded, nothing more.
The four grades and what counts as proof
| Grade | What it needs | Example |
|---|---|---|
| Proven | A primary page from the engine owner, or a measured study with a named sample and method that we opened ourselves. | Google's spam policies list excessive link exchanges as link spam. |
| Claimed | Anything less: a vendor blog, a case study with no control, a correlation, a conference quote reported second hand. | "Titles that echo ChatGPT's hidden searches get cited more" is a correlation. Useful, still claimed. |
| Myth | Contradicted by a better source. | "llms.txt gets you into Google's AI answers." Google says Search does not use it. |
| Not applicable | No search claim at all. | A clever prompt trick with nothing to do with visibility. |
Two rules keep the grading straight. A vendor blog is "claimed" at best, even when we like the vendor, because vendors sell the tool that measures the problem. And a correlation stays "claimed" even when the sample is huge. Seventeen million rows can still describe something that is not a cause.
A wrong number that sounds sourced does more damage than no number. We once caught a widely shared "lift" figure that was really an age gap between two groups of pages. Since then, every figure goes back to its source before it goes anywhere near a client.
The ten-minute method
- Write down the claim in one sentence. Strip the hype. "Do X and Y happens." If the post makes three claims, grade three.
- Find the primary source. The engine's own documentation, or the study itself, not the article about it. Note its date. If the newest source is two years old, say so.
- Check the rules. Google's spam policies (last updated 28 August 2026), the review rules of the platforms you use, and consumer law where you sell. Anything that touches reviews gets the extra check below.
- Check whether we already do it. Half of what goes viral is an existing practice renamed. If it is in our method already, the answer is "already present", and the only question is whether the post adds a detail worth keeping.
- Decide. Adopt (do it as described), adapt (do the sound part, drop the rest), already present, or reject.
- Write the measurement before acting. One row: what changes, which metric, the baseline, the target and the date we check. No row, no test.
- Get a second opinion before the method changes. Whoever graded the tip does not approve the change to our method. A second reviewer checks the grade against the sources.
Most tips take less than ten minutes. The ones that take longer usually deserve it, because they touch rules or money.
Eleven tactics we graded
These are real tips we triaged this year, with the verdict we reached and why. The grade is about the evidence. The decision is about what we do with it.
| Tactic | Grade and decision | Why |
|---|---|---|
| "ChatGPT's hidden searches decide what gets cited" | Depends Claimed. Adapt. | Google defines fan-out in its own guide. One small study found about 40% of ChatGPT's cited sources ranked for the hidden queries. We harvest those words and use them on existing pages. See query fan-out. |
| Filter Search Console to queries of four or more words | Works Proven. Adopted. | A regular expression filter in Search Console does this. We run it monthly; on our own site it surfaces full questions like "does claude work in thailand" (our Search Console). See long-tail mining. |
| "Those long queries are AI fan-out searches" | Hype Claimed. Not used as a reason. | Plausible, unproven. Seer measured Gemini 3 fan-out queries averaging 6.7 words, which fits, but it is not proof that your long queries come from AI. They are worth working on either way, because they are real impressions. |
| Recover lost backlinks | Works Claimed. Adopted. | No study needed: a link that broke because a page moved is fixed with a redirect or a polite note. Low risk, low cost. |
| Turn unlinked brand mentions into links | Works Claimed. Adopted. | A normal editorial request. We think brand mentions matter at least as much as the link for AI answers, so we track both. |
| Build a free tool to earn links | Depends Claimed. Adapt. | Only when it does a real job for a buyer with data the business owns. Never distributed as a widget with keyword links, which the spam policy names. |
| Get into "best of" lists | Depends Claimed. Adapt. | Local SEO experts rank "presence on expert curated best-of lists" first for AI search in Whitespark's 2026 survey. That is expert opinion, not measurement. Earned inclusion only. Paid placements must be marked sponsored or nofollow. |
| Three-way link exchanges (A to B to C to A) | Against the rules Myth. Rejected. | Google lists excessive link exchanges as link spam. Adding a third site does not change the purpose. |
| Publish a statistics page | Depends Claimed. Adapt. | Only with first-party numbers. The GEO paper found statistics helped on a 2023 research engine. Copying other people's statistics adds nothing. |
| Add an llms.txt file | Skip it Myth for search. Rejected as a lever. | Google Search does not use it. SE Ranking found no link to citations; Ahrefs found 97% of files got zero requests in May 2026. In our Bangkok study it was less common among the most named businesses (5 of 28) than the rest (26 of 82). |
| Buy reviews, or ask only happy customers | Against the rules Rejected. | Against Google Business Profile policy and against the law in the US and UK. See below. |
How to read the study behind a tip
Most tips lean on a study, and most studies are fine. The problem is the sentence that travels without them. Before we accept a number we ask six questions.
| Question | Why it matters | Real example |
|---|---|---|
| Who measured it? | A vendor measuring the problem its tool solves is not neutral, even when honest. | Many AI visibility studies come from companies that sell AI visibility tracking. |
| What was the sample? | Ten sites and ten million pages say different things. | A single site reporting AI visitors converting 23 times better is one site, and its author says so. |
| Did the method change? | Two editions of one study can disagree because the counting changed. | Ahrefs' March 2026 update put AI Overview citations from the top 10 at 37.9%, down from 76% in July 2025, after a change in how citations were parsed. |
| Is it a correlation? | Pages that do X may be cited more because they are better pages, not because of X. | Cited pages are about three times more likely to carry schema, yet adding schema did not move citations in a matched test. |
| When was the data collected? | AI search changed every few months in 2025 and 2026. | A 2024 crawler study can still be right, but it needs a date next to it. |
| Does the window contain a Google update? | A before and after across a core update measures Google. | Eight ranking updates are listed on the dashboard between December 2025 and September 2026. |
None of this makes a study useless. It tells you how much weight it can carry. A correlation can point us to a test. Only the test, on our own pages, earns a change to how we work.
A worked example: the long-tail reel
In October 2026 a short video went round claiming that the long queries in Search Console are really the hidden searches AI assistants run, and that you should mine them and add sections to your pages. This is how it went through the method.
- Claims. There were two. One: a filter can isolate queries of four or more words. Two: those queries are AI fan-out searches.
- Sources. The filter is a documented Search Console feature. For the second claim we found no study at all. The nearest evidence, Seer's research on Gemini 3, shows fan-out queries average about seven words and mostly have no measurable search volume. That fits the idea but does not prove it.
- Rules. Nothing in it breaks a policy, as long as the answers go into existing pages. Writing a new page for every long query would be the scaled content pattern Google warns about.
- Already doing it? Partly. We already used long questions as section headings. We did not have a monthly step for it.
- Decision. Adapt. Keep the mining, drop the reason. Long queries are real impressions on your own site at a position you can move. That is reason enough, whoever typed them.
- Measurement. Each section added gets a row: the query, its page, its position and clicks, checked four weeks after shipping.
The result is now a monthly step in our engine, with its own tests, and a guide: mine your long-tail searches. The tip was useful. Its explanation was not needed.
The tactics that break the law, not just the rules
Most bad SEO tactics risk a ranking. Review tactics risk a fine. This is where we are strictest, because the tip-givers almost never mention it.
| Rule | What it says | Date |
|---|---|---|
| US FTC rule on reviews and testimonials | Bans fake or AI-generated reviews, compensation that depends on a positive review, undisclosed insider reviews and buying fake social indicators. | Final rule announced 14 August 2024, in effect since 21 October 2024. |
| UK Digital Markets, Competition and Consumers Act 2024, Schedule 20 | Commissioning fake reviews or hiding incentives is banned, and traders must take reasonable steps to prevent and remove them. | Act of 2024, in force since 6 April 2025. |
| Google Business Profile policy | No paid or incentivised reviews, no review gating (asking only happy customers), no pressure on the premises. | Current policy page, undated. |
| Google review snippet guidelines | No fake or undisclosed incentivised reviews on the page or in markup. | Rule added in 2026; page updated 8 September 2026. |
What about a small discount for anyone who leaves a review, good or bad?
Google Business Profile says merchants may not offer incentives for reviews, full stop. So no. Ask every customer, in the same way, at a natural moment, and make it easy. That is allowed by Google, the FTC rule and the UK DMCC Act. More on this in local and map visibility.
How we measure a tactic we adopt
A tactic is not adopted until it has a measurement row. This is the shape we use, with an example.
| Field | Example |
|---|---|
| Change | Add the buyer's exact question as an H2 with a two-sentence answer on the pricing guide. |
| Metric | Search Console clicks for that page, plus position for the exact query. |
| Baseline | The last 28 days before the change, with dates. |
| Target | A specific number, not "improve". |
| Check date | 28 days after the change is indexed. |
| Confounders | Any Google update in the window, listed on the Search Status Dashboard. If one overlaps, the result is inconclusive unless the effect is clearly bigger than the usual swing. |
At the check date the row gets one of four results: win, loss, flat or inconclusive. Most real results are flat or inconclusive. That is not failure. It is how you avoid building a strategy on luck.
For AI visibility tactics the bar is higher, because answers vary so much run to run. We explain that maths in measuring AI visibility.
Red flags, mistakes and the checklist
Red flags
- One screenshot of one AI answer as proof.
- "Hack", "secret", "Google doesn't want you to know".
- A score out of 100 with no method.
- "A study found" with no study linked.
- A before and after chart with no dates, or a window that contains a Google update.
- The poster sells the tool that measures the problem.
- No mention of rules, reviews or consumer law in a tactic about reviews or links.
Mistakes
- Trying five tips at once on one page, so nothing can be attributed.
- Adopting a tactic for "all clients" when it only fits one kind of business.
- Grading your own idea. The person who proposes a change should not approve it.
- Treating a correlation as a lever.
- Letting a tip skip the measurement row because it "obviously works".
Checklist
- Claim in one sentence. No hype words left.
- Primary source found and dated. Not the article about it.
- Grade given. Proven, claimed, myth or not applicable.
- Rules checked. Spam policies, platform review rules, consumer law.
- Already doing it? If yes, keep any useful detail and stop.
- Decision. Adopt, adapt, already present or reject.
- Measurement row written. Metric, baseline, target, check date, confounders.
- Second reviewer signed off. Before anything in the method changes.
This is the last guide in the track. Go back to the start, what works in search and AI visibility in 2026, or send us the tip you are unsure about through the AI visibility diagnostic.
Sources and dates
Every external fact in this guide was read on 20 October 2026. Each source shows its publisher and the date it was published or last updated, and we flag what changed in 2026. Studies are their authors' measurements, usually correlations, not promises of results. Our own numbers come from our own Search Console and tools and say so where they appear.
- Google's Guide to Optimizing for Generative AI Features on Google SearchGoogle Search Central · Last updated 10 July 2026 · Read 20 October 2026. Changed in 2026: Guide added 2026-05-15 (changelog); llms.txt clarification note added 2026-06-15; page last updated 2026-07-10.
- Prohibited & restricted content - Google Business Profile HelpGoogle Business Profile Help · Undated page · Read 20 October 2026.
- Review snippet (Review, AggregateRating) structured dataGoogle Search Central · Last updated 8 September 2026 · Read 20 October 2026. Changed in 2026: Guideline on fake and undisclosed incentivized reviews added 2026-07-24 (changelog).
- Spam policies for Google web searchGoogle Search Central · Last updated 28 August 2026 · Read 20 October 2026. Changed in 2026: Site reputation policy section updated 2026-08-28 for the EEA; 2026-04-13 back button hijacking added to malicious practices; 2026-05-15 clarification that spam policies apply to generative AI responses.
- Google Search Status Dashboard: ranking incidentsGoogle Search Status Dashboard · Undated page · Read 20 October 2026. Changed in 2026: Listing shown on read; older 2025 incidents may exist beyond the summary view.
- dlvx.co Search Console performance, 19 July to 16 October 2026DLVX (own Search Console, dlvx.co) · Our own measurement, data to 16 October 2026 · Read 20 October 2026.
- Who AI recommends in Bangkok, 2026DLVX · Published 3 October 2026 · Read 20 October 2026.
- FTC Announces Final Rule Banning Fake Reviews and Testimonials (press release)FTC · Published 14 August 2024 · Read 20 October 2026.
- The Consumer Reviews and Testimonials Rule: Questions and AnswersFTC · Last updated November 2024 · Read 20 October 2026.
- Digital Markets, Competition and Consumers Act 2024, Schedule 20UK Government (legislation.gov.uk) · Undated page · Read 20 October 2026.
- Does AI Search Traffic Convert Better Than Traditional Search? For Ahrefs, Yes: 0.5% of Visitors Drove 12.1% of SignupsAhrefs · Published 16 June 2025 · Read 20 October 2026.
- Update: 38% of AI Overview Citations Pull From The Top 10Ahrefs · Published 2 March 2026 · Read 20 October 2026.
- We Analyzed 137K Sites: 97% of llms.txt Files Never Get ReadAhrefs · Published 15 June 2026 · Read 20 October 2026.
- We Tracked 1,885 Pages Adding Schema. AI Citations Barely Moved.Ahrefs · Published 11 May 2026 · Read 20 October 2026.
- Only 40% of ChatGPT Sources Come from Google & Bing for Known Fan-Out QueriesGrow and Convert · Last updated 30 March 2026 · Read 20 October 2026.
- GEO: Generative Engine OptimizationPrinceton / Georgia Tech / IIT Delhi / Allen AI authors (arXiv) · Published 16 November 2023 · Read 20 October 2026.
- Initial Research: Gemini 3 Query Fan-OutsSeer Interactive · Published 21 November 2025 · Read 20 October 2026.
- Does LLMs.txt impact your AI visibility and citations? No, according to researchSE Ranking · Published 15 October 2026 · Read 20 October 2026.
- Local Search Ranking Factors 2026Whitespark · Published 6 November 2025 · Read 20 October 2026.
- The Digital Markets, Competition and Consumers Act 2024 (Commencement No. 2) Regulations 2025, regulation 2legislation.gov.uk · Undated page · Read 20 October 2026.
Want this run on your own site? Start with the AI visibility diagnostic, or see the whole Get Found track.