Disclosure. Entry eight, Jacob Whitmore, owns Blogging Titan. He is in this roundup for a test he ran and published on his own site that answers the question directly. His entry is held to the same rule as the others: his published words, linked, with our data beside them. None of the other seven were contacted, paid or asked for a quote. Everything below was already public.

How we built this roundup

Most expert roundups are made by emailing fifty people one question and pasting the replies. Lily Ray, entry three, has written about where self-promotional list formats drift. We did it the slow way. Each entry is something the person had already published, read on its original page, quoted exactly, dated and linked. Where the source is a report about someone’s research and not their own writing, we say so.

After each quote comes what our own studies found on the same question, including the places where our numbers push back.

The eight experts

They are listed in the order the argument builds, starting with Google itself.

1. Danny Sullivan, Google

Sullivan, formerly Google’s search liaison and now a director within Google Search, gave the keynote at WordCamp US in August 2025. Search Engine Land published a transcript of the part about AI optimisation.

Good SEO is good GEO, or AEO, AIO, LLM SEO, or LMNOPO. So, they’re all fine. What I’m trying to say is don’t panic.

Danny Sullivan, WordCamp US keynote, 28 August 2025, via Search Engine Land

What our data says. Mostly yes, with one gap that SEO cannot close. When we asked ChatGPT 100 blogging questions, it cited nobody at all on 48 of them. The silence clustered on stable how-to advice, while it opened the web for 15 of 15 tool and software questions. Good SEO cannot win a citation on an answer the model writes from memory. The full breakdown is in our AI citation league table.

2. Aleyda Solis, AI Search Optimization Checklist

Solis keeps a public AI search optimisation checklist, last updated in September 2026. Near the top, before any of the steps, she sets one limit.

Prompt injection, fabricated reviews, manufactured community signals and citation gaming might end up producing short-term improvements, but they will get eventually addressed, and will trade durable trust for visibility that can backfire.

Aleyda Solis, The AI Search Optimization Checklist, updated September 2026

What our data says. The short term may be shorter than it sounds. In our citation lottery, three of 20 Google AI Overviews rewrote most of their sources inside a single hour with nobody touching anything. A citation won by gaming has to survive those re-rolls too, and the reworded questions in entry six are far harsher on it.

3. Lily Ray, on her Substack

In February 2026 Ray wrote about companies publishing “best of” listicles on their own blogs with themselves in the top spot, and about competitors listing one another in turn. She called the second habit

a modern twist on reciprocal linking: you mention me, and I’ll mention you.

Lily Ray, Is Google Finally Cracking Down on Self-Promotional Listicles?, 3 February 2026

What our data says. This one applies to us, so we will say it plainly. A roundup is close to the format her point is about, and this one includes the owner of the site that publishes it. That is why the disclosure sits at the top, why nobody here was asked to return a favour, and why every entry links to its source. The habit is common. In our Stake Test, 32 of the 40 page-one pages for “best blogging tool” searches either earned a commission on their picks or came from a vendor recommending its own product.

4. Kevin Indig, Growth Memo

This entry reports Indig’s research through Search Engine Land, which covered his analysis of 1.2 million AI answers and 18,012 verified citations in February 2026. It described a consistent “ski ramp” pattern in where ChatGPT takes its citations from on a page. 44.2 percent came from the first 30 percent of the content, 31.1 percent from the middle, and 24.7 percent from the final third (Search Engine Land, 18 February 2026).

What our data says. Front-loading only helps if the top of the page holds something worth quoting. In our census of 161 page-one blogging pages, 71.2 percent of the statistics had no source beside them. Put the specific, sourced claim in the first third, with the link right next to it.

5. Rand Fishkin, SparkToro

In January 2026 SparkToro published research run with Patrick O’Donnell of Gumshoe.ai. About 600 volunteers ran 12 prompts through ChatGPT, Claude and Google’s AI in November and December 2025, producing 2,961 responses, and the study asked whether the brand recommendations that came back were consistent enough to measure.

If you ask an AI tool for brand/product recommendations a hundred times nearly every response will be unique

SparkToro, 27 January 2026

The post goes on to list the three ways they differ: which brands appear, in what order, and how many.

What our data says. Our numbers look calmer, and the gap is useful. When we pulled the same 20 AI Overviews three times in an hour, 17 kept an identical source list. That was one browser, one location and one hour. SparkToro’s prompts were identical too, but they were run hundreds of times by different people on their own devices, and they asked for brand lists, which are not the same thing as cited sources. Read together, the two studies suggest that identical wording buys stability only in a narrow window, and that across many runs even the same words come back different.

6. Amanda Natividad, SparkToro

Natividad asked eight friends with similar lives to find a local basketball league for their children, and watched how differently they prompted AI. She tied it back to the SparkToro study, where the human-written prompts had an average semantic similarity of 0.081.

In other words: people don’t “search” anymore. They ask. And increasingly with LLMs, they… talk.

Amanda Natividad, Why Everyone Prompts AI Differently, 11 February 2026

What our data says. We measured what that does to citations. We asked Google the same question three different ways, and only 14 percent of its sources survived all three phrasings. In our earlier lottery test, where the question was asked identically three times, 87.5 percent survived. When no two readers type the same question, a position for any one wording is worth very little.

7. Mike King, iPullRank

In a webinar on Google’s AI Mode with Andrea Volpini of WordLift, recapped by iPullRank in July 2025, King pointed at the mechanism behind a lot of this.

The query fan-out technique…is the big thing that I think we need to be thinking about in our space.

Mike King, iPullRank AI Mode webinar recap, July 2025

Fan-out is the system splitting one question into many sub-searches and writing the answer from what those return.

What our data says. Fan-out is a likely explanation for the 14 percent figure in entry six: reword the question and the sub-searches change, so the sources change with them. We have not measured Google’s fan-out directly, so treat that as a probable mechanism and not a finding. The practical response for a blog is to cover the sub-questions a topic splits into, on the page or across linked pages, so you are in the pool whichever way the question fans out.

8. Jacob Whitmore, jacobwhitmore.com

Whitmore has worked in SEO for around 12 years. At jacobwhitmore.com he publishes dated research on AI search with the raw data attached. He also keeps a public ledger of predictions, each to be scored right or wrong on a stated date, with misses kept visible. On 9 October 2026 he asked ChatGPT and Gemini for the best UK accounting software in three ways. The first prompt gave no persona, the second posed as a self-employed plumber, and the third as the finance director of a manufacturer with 250 staff. That made 12 answers in all.

There’s no single top three to rank in. The shortlist is built for the person in the prompt.

Jacob Whitmore, Does ChatGPT recommend different brands to different customers?, 9 October 2026

The plumber got much the same list as the no-persona prompt. The finance director got a different market: three of the four finance-director answers recommended none of the products the no-persona answers had named.

What our data says. This is the person-shaped half of the asker effect. Our studies changed the wording and his changed the asker, and both moved the answer. Twelve answers on two engines is a small test, which his write-up lays out in full, so read it as a direction to check in your own niche.

Where the eight agree

Read side by side, the entries reduce to three points. Sullivan says the fundamentals of SEO carry over, and Solis and Ray warn against gaming the systems. Indig’s data says to put the substance where the machine reads first. Fishkin, Natividad, King and Whitmore all describe answers that will not hold still: they vary between runs, between wordings, between the sub-searches a question fans into, and between the people asking.

That last point is the one most AI visibility advice still skips, so we gave it a name.

The asker effect

The asker effect is our name for a pattern that runs through our own studies and through two of the entries above. Over a short window, AI answers are fairly stable against the clock and unstable against the asker. Ask Google the identical question again within the hour and it mostly repeats itself, as 17 of 20 Overviews did in our lottery. Change the words, as Natividad’s friends naturally did, and in our paraphrase test only 14 percent of the sources survived. Change the person, as Whitmore’s personas did, and the shortlist changed for the finance director. SparkToro’s data adds a caution on top: across hundreds of runs, even identical prompts rarely return the same list.

It changes what ranking means in AI search. There is no fixed position to hold. There is a share of the answers given to many different askers. Search Engine Land’s write-up of the SparkToro data makes the same point: how often a brand appeared held up as a measure, while exact rank did not.

What we would do on a small blog

Six changes, all of them things a one-person blog can do this month.

1. Write one page for each real sub-question a topic splits into, and link the pages together, so you sit in the pool for more of the fan-out.
2. Put your specific, sourced claim in the first third of the page with the link beside it.
3. Publish something only you have, such as your own count, test or dataset. Half the entries above rest on data their author collected, and that is what makes them quotable.
4. Track presence across phrasings and personas. Run each target question three ways and as two or three kinds of reader, and record whether you appear at all.
5. Do not let pages rot. In our freshness study, pages left untouched for two years or more turned up nearly twice as often among the uncited pages as among the cited ones.
6. Leave the gaming alone, for the reason Solis gives in entry two.

Related guides

• We asked Google the same question three ways. Only 14% of its sources survived
• The citation lottery: the same 20 AI Overviews, three times in one hour
• The AI citation league table: what ChatGPT and Perplexity cite
• Who wrote the 951 pages ChatGPT and Perplexity cited
• The page-one blogging census

Method and limits

Every source was opened on 17 October 2026. Each quote was copied from the page text with its punctuation intact. That includes the ellipsis in Mike King’s line and the missing full stop at the end of the SparkToro sentence, which continues into a list on the original page. Sullivan’s words come from Search Engine Land’s transcript of a talk. Indig’s entry reports his findings through Search Engine Land and quotes none of his own sentences. The SparkToro study size and dates are from SparkToro’s own post, linked in entry five. The Blogging Titan figures come from the studies linked in each entry, where the sample sizes, dates and caveats are set out. The selection is ours: we chose people who have published specifically on AI citation, and a different eight would read differently.