Part seven
Myths and facts: thirty one marketing claims checked against the source
Every claim in this part was checked against the primary source before it went on the page. What follows is the register: what failed, what survived, and the four ways a marketing statistic goes wrong.
-
2 in 3
Of the widely repeated platform claims we checked did not survive contact with the primary source
This document 1
-
2.2 not 21
Reddit's share of all Google AI Overviews citations. The famous 21 percent is a top ten share, printed in the same study
Profound, June 2025 2
-
2014
The date of the newest large-scale measurement of dark social with a stated survey sample
RadiumOne 3
-
~500
Marketers self-reporting their own returns in the survey behind "email returns $36 for every $1"
Litmus, July 2025 4
-
Zero
Studies found testing whether founder-led content outperforms brand-account content, in either direction
Searched 8 Aug 2026 5
-
Wrong denominator
The number is real and is a share of something much smaller than you assumed. The most common failure in this document.
-
Right once
The number was true on the date it was measured and the thing it measured has since been replaced. X's ranking weights are three years old.
-
No source at all
Confident, specific, widely repeated, and it traces to nothing. The $4,800 retainer, the 9 to 1 rule, three platforms before deciding.
-
Sold by the person measuring
The company publishing the study also sells the service the study recommends. True of almost every AI visibility number in circulation.
- Counted across the eight platform chapters in Part Two. n/a
- Profound, AI Platform Citation Patterns , 5 June 2025. 2nd
- RadiumOne, press release 1 December 2014. 2nd
- Litmus, The ROI of Email Marketing , 16 July 2025. 2nd
- Searched 8 August 2026, across academic, platform and vendor sources. n/a
The four ways a marketing statistic goes wrong.
Across every chapter in this document the failures fell into the same four shapes. Once you can name them you can check a claim in about ninety seconds, and you will not need this document to do it.
-
01
The denominator is not what you assumed
The number is real and correctly reported. It is a share of something much narrower than the sentence implies. "Reddit is 21 percent of Google AI Overviews" is 21 percent of the top ten sources, and the same study prints 2.2 percent of all citations two lines away. "Eighty-four percent of sharing is dark social" is a share of sharing that started from one company's own share buttons. 6
-
02
It was true once
The measurement was sound on the day it was taken and the thing it measured has since been replaced. X's famous ranking weights are from 5 April 2023 and the code that used them has been rewritten twice. The segmentation statistic still quoted this year was published in February 2017. 7
-
03
There is no source
Confident, specific, attributed to a credible-sounding institution, and not in the document it is attributed to. The $4,800 agency retainer is not in the report it is credited to. The Reddit 9 to 1 rule is not in Reddit's policy. "The average consumer uses three platforms before deciding" is credited to nobody at all. 8
-
04
The measurer is the seller
The company publishing the study also sells the service the study concludes you need. That does not make it wrong. It does mean the sample, the metric and the conclusion were all chosen by an interested party. Every AI visibility statistic in circulation comes from a company selling AI visibility software. 9
- Profound, 5 June 2025; RadiumOne report page, undated. 2nd
- GitHub, X ranking weights, 5 April 2023; Mailchimp, 1 February 2017. 1st/2nd
- AgencyAnalytics 2026 report; Reddit Help Centre, 19 May 2026; SOCi, 29 July 2026. 1st/2nd
- Profound, Ahrefs, Semrush, Similarweb, Otterly, BrightLocal, Litmus and others, throughout this document. 2nd
The platform register, part one.
Sixteen claims about how the platforms work, checked against what the platform itself publishes. Fourteen of them fail. These are the first eight.
| The claim | Verdict | What the platform actually publishes |
|---|---|---|
| Sends per reach is an Instagram metric | Not a metric | No field called sends exists in Meta's documentation. The metric is shares. Instagram's own dated ranking documentation does not mention sends at all; the doctrine comes from informal video posts |
| Instagram penalises engagement bait | Wrong platform | Engagement bait is a Facebook content distribution guideline. The words appear nowhere in Instagram's Recommendation Guidelines |
| Instagram says follower count is irrelevant | Overstated | Instagram says distribution is "not just" follower count, and separately lists number of followers as a popularity signal |
| TikTok shows a video to 300 to 500 people first | No source at all | TikTok describes per-viewer ranking from "a large collection of eligible content." The circulating numbers do not even agree with each other |
| Completion rate is TikTok's key metric | Not a TikTok metric | TikTok has never named a most important metric and has no metric called completion rate. It shows "watched the full video" |
| Posting more than once a day hurts TikTok reach | Contradicted | TikTok's own guidance says "aim to post 1-4 times per day." An 11.4 million post panel shows median views falling across the whole cadence range, which is dilution, not a penalty |
| Facebook organic reach is about 2 percent | Never published by Meta | Meta has published no organic reach figure for Pages, ever. Every circulating percentage is a tool vendor's own panel |
| Facebook and TikTok demote AI-generated content | No source | Meta labels rather than demotes: "We will keep this content on our platforms so we can add informational labels and context." The word AI appears nowhere in TikTok's ineligibility list |
Every verdict above is sourced in full in the chapter it comes from, in Part Two. Platform documentation was retrieved between 28 July and 8 August 2026, and where a page carries no date, that is stated at the point of use.
The platform register, part two.
Continued. LinkedIn, X and Reddit, which between them carry the most confidently repeated and least sourced claims in the whole document.
| The claim | Verdict | What the platform actually publishes |
|---|---|---|
| Posting in a Facebook Group reaches non-members | Unsupported | No Meta page states Group content is recommended to non-members |
| LinkedIn penalises posts with links | Denied on record | LinkedIn has never published anything saying links are penalised, and a named LinkedIn product director says on record they are not |
| X's published weights are current | April 2023 | X's own file says they "can be adjusted at any time." In the September 2025 release all 29 weight defaults are zeroed. In the May 2026 release the module defining them is absent from the repository |
| X Premium gives 4x reach | Not in the code | The multipliers existed in 2023 and do not appear anywhere in the September 2025 release. X announced no removal |
| Links get throttled on X | Owner only, self-contradicted | X the company has published nothing. Its owner said links reduce reach in 2023 and 2024, then in April 2025 that "there is no explicit rule limiting the reach of links in posts" |
| Reddit's 9 to 1 self-promotion rule | Not Reddit policy | Reddit's spam guidance, updated 19 May 2026, contains no ratio of any kind. The published rule is behavioural: repetition, volume, unsolicited contact, account manipulation |
| Most subreddits ban links | No dataset exists | No published measurement of link restrictions across subreddits was found, and Reddit publishes no site-wide karma or account age threshold |
| Reddit is 21 percent of Google AI Overviews | Misquoted | 2.2 percent of all AI Overviews citations. The 21 percent is a share of the top ten sources only, and both numbers are in the same study |
Every verdict above is sourced in full in the chapter it comes from, in Part Two. Platform documentation was retrieved between 28 July and 8 August 2026, and where a page carries no date, that is stated at the point of use.
The search register.
Eight claims about search and AI, from the part of marketing where the most money is currently being spent on advice and the evidence is thinnest.
| The claim | Verdict | What the source actually says |
|---|---|---|
| You need an llms.txt file | Not for Google | "You don't need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search (including its generative AI capabilities), as Google Search itself doesn't use them" |
| You need special schema for AI search | It does not exist | "Structured data isn't required for generative AI search, and there's no special schema.org markup you need to add" |
| GEO improves visibility by up to 40 percent | On a machine the authors built | The paper fetched "top 5 sources from Google search engine" and generated answers with "gpt3.5-turbo," on a 10,000 query benchmark. Not Google, not AI Mode, not ChatGPT. The preprint is 33 months old |
| E-E-A-T is a ranking factor | A description, not a lever | "No single rating can directly impact how a particular webpage, website, or result appears in Google Search, nor can it cause specific webpages, websites, or results to move up or down" |
| AI answers appear on about half of searches | Half of tracked keywords | 48 percent is one vendor's tracked keyword set, whose size has never been published. Real searches by real people run at 18 to 20 percent on two separate panels |
| Posting on your Business Profile helps you rank | Not on the one test that exists | 441 keywords per location across three listings, weekly posts for nine weeks, no measurable local pack effect. The firm that ran it reassigns posts to conversion |
| Blocking AI training has no downside | Depends who, and on Cloudflare | OpenAI, Anthropic, Amazon and Perplexity separate training from search. Google, Apple and Meta do not. And from 15 September 2026 Cloudflare blocks Googlebot for anyone who selected to block Training, including via the legacy one-click toggle |
| Google will start booking your restaurant tables | Not what was announced | "For select categories like home repair, beauty or pet care, you can ask Google to call businesses on your behalf." US only. Restaurants are not named anywhere in the announcement |
Sourced in full in Parts Three, Four and Five. Google Search Central documentation retrieved 8 August 2026; academic citations given with journal and date at the point of use.
The email and money register.
Seven more, from email and from the places where the invoices are largest.
| The claim | Verdict | What the source actually says |
|---|---|---|
| Email returns $36 for every $1 | A survey of marketers | Roughly 500 marketing professionals estimating their own return. Only 30 percent of them said $36 to $50, and 30 percent of respondents are missing from the three bands quoted |
| A good open rate is around 30 percent | Measures a proxy server | 55 to 60 percent of recorded opens are affected by Apple's proxy and are "not considered reliable opens." A separate dataset puts the true rate near 11 percent |
| You must get DMARC to enforcement | Google says otherwise | "Your DMARC enforcement policy can be set to none." Yahoo asks for at least p=none. The providers require the record, not the enforcement |
| Personalised subject lines and emoji lift opens | Associated with fewer | Across 4.4 billion messages: 35.78 percent personalised against 41.87 not, and 37.5 percent with an emoji against 42.23 without |
| Bidding on your own brand name protects you | Measured, and it did not | "We show that brand-keyword ads have no measurable short-term benefits." Average returns across the experiment were negative |
| The average agency retainer is $4,800 a month | Not in the report | 494 respondents, and no dollar figures anywhere in the document it is attributed to. Earlier editions all return 404 |
| Trust in advertising is collapsing | Nobody measures it | The study normally cited measures institutional and interpersonal trust across 33,938 respondents in 28 countries. It does not measure trust in advertising, and we found nothing that does |
Sourced in full in Parts Three, Four and Five. Google Search Central documentation retrieved 8 August 2026; academic citations given with journal and date at the point of use.
What survived.
A document that only knocks things down is not worth reading. These are the claims that held up against a primary source, and they are the ones worth building on.
| The claim | Verdict | The evidence, and its limits |
|---|---|---|
| Most of what people see is no longer from accounts they follow | Holds | Meta's own figures, 30 percent of Facebook feed and over 50 percent of Instagram content, April 2024. Never restated since, so use the direction and Meta's own sentence rather than the numbers 10 |
| People are exploring far fewer websites | Regulator-measured | The proportion of UK adults exploring new websites fell from 70 percent to 56 percent in one year, and active posting among social media users fell from 61 to 49. 7,533 adults, fieldwork September to November 2025 11 |
| Two thirds of searches end without a click | US and UK, 2026 | 68.01 percent US and 69.5 percent UK, January to April 2026, up from 60.45 percent in 2024. Clickstream panel, size not published. No global figure exists 12 |
| AI answers are grounded in ordinary search ranking | Google's own words | Grounding relies "on our core Search ranking systems to retrieve relevant, up-to-date web pages from our Search index," and "the best practices for SEO continue to be relevant" 13 |
| These systems cite passages, not pages | Three independent lines of evidence | Google appends text fragments to citation links, Anthropic's citation payload caps quoted text at 150 characters, and an academic study of 21,143 citations found the pages that get used contain "definitions, numerical facts, comparisons, and procedural steps" 14 |
| Reviews move revenue for independent restaurants | Tax records | A one-star increase produced a 5 to 9 percent revenue increase across 3,582 restaurants, from state tax filings, "driven by independent restaurants; ratings do not affect restaurants with chain affiliation." Data ends 2009 15 |
| Triggered email outperforms broadcast email | Confounded by design | $3.41 per automated email against $0.155 per campaign email, 2025 panel. Automated mail fires at a moment of intent to a self-selected person, so the comparison is not like for like. It also gets unsubscribed from three times as often 16 |
| Being easy to think of and easy to find is what grows a business | Strong, and not built on restaurants | Mental and physical availability come from a large replicated body of work built overwhelmingly on repeat-purchase consumer goods. The idea travels well to a restaurant. The evidence base was not built on one 17 |
- Meta, Q1 2024 and Q3 2024 earnings calls. 1st
- Ofcom, 2 April 2026. 1st
- SparkToro with Similarweb, June 2026. 2nd
- Google Search Central, 10 July 2026. 1st
- Google, Anthropic and arXiv 2604.25707, April 2026. 1st
- Luca, Harvard Business School. 1st
- Omnisend, 12 May 2026, 2025 data. 2nd
- Ehrenberg-Bass Institute; Sharp, 2010. 1st
And the questions nobody has answered.
These are the gaps we hit repeatedly. Each one is a question a business owner reasonably asks and nobody has published an answer to. They are listed here so that when somebody gives you a confident number on one of them, you know where it came from.
- How much sharing actually happens privately. The category's headline statistics are between eight and fourteen years old, from companies selling share buttons. The newest study with a stated method is a ten-day test with about a hundred people, from April 2023
- Whether founder-led content outperforms brand-account content. Not a controlled study, not a platform disclosure, not an academic paper. Nothing in either direction
- How audiences respond to visibly AI-generated content. No published measurement with a stated sample
- Whether asking customers how they heard about you produces accurate answers. Recommended universally, tested by nobody. Two separate research passes came back empty
- How anybody actually chooses a restaurant. How many places they consider, which surfaces they check, how long it takes. With a stated sample, in the US or UK, recently. We could not find one
- Whether local sponsorship works. Sports teams, school programmes, local events. No study, no survey, no case series with a method
- Where AI assistants get their local recommendations. Maps, review sites, directories, something else. No first-party statement from any of them
- What proportion of sites now block AI crawlers. The two obvious sources are dashboards whose numbers do not survive being fetched
Each gap above was searched across academic, platform, regulator and vendor sources between 28 July and 8 August 2026, and each is stated at the point in the document where it arises. Where a page could not be retrieved, that is recorded separately from an absence of evidence, because they are different things.