LLM SEO: the levers we tested on sites we manage

The LLM SEO guide we read before writing this ends with a familiar list of levers: write for Google and Bing, add schema and short answers, refresh your dates, earn mentions. We measured those levers against our own readings on wiserstaff.com, my own blog of HR-software reviews published under a pen name. The readings cover Microsoft Copilot citations from 16 August to 3 October 2026, Google positions for September, and fixed checks of ChatGPT and Google’s AI Mode on 4 October. This page scores each lever from the sites we manage: wiserstaff.com for every verdict, and riin-art.com, a client’s shop, for one Google check.

What LLM SEO means

LLM SEO is the work of getting a site cited in AI answers, such as those from ChatGPT, Copilot or Google’s AI Overviews. Classic SEO aims at a ranked link; LLM SEO aims at being named as a source. Each assistant picks sources its own way, so a lever that works for one may do nothing for another.

The field has several names for the same goal: generative engine optimisation (GEO), answer engine optimisation (AEO), AI search optimisation. The name matters less than the measurement behind it. We read citations engine by engine, the method our AI Citation Meter is built on, and we never add one engine’s count to another’s. What being cited was worth in visits and money is a separate question, covered in what our citations turned into.

LLM SEO vs SEO: what changed on our sites

On our sites, the difference between LLM SEO and SEO depended on the engine. In our small Google checks, Google’s AI surfaces cited pages that already ranked in Google; Microsoft Copilot cited pages that Google ranked poorly. For Google, SEO and LLM SEO pulled the same way; for Copilot, they did not.

The two cases that show it best:

• Google followed its own ranking. On riin-art.com, a shop I built for a client, whose figures I publish with the client’s agreement, Google’s AI Overview cited the site on all 6 questions where it already appeared in the organic results, and on none of the 3 where it did not (9 readable answers, French and Arabic questions, 22 September 2026). On wiserstaff.com, the one Google AI Mode citation in our ten fixed questions of 4 October went to a comparison page that averaged position 5.1 on Google in September.
• Copilot did not. On wiserstaff.com, the 19 content pages in Copilot’s list of our most cited pages that had a Google position averaged 39.5 on Google in September; the 22 outside that list averaged 26.5. The one answer in which ChatGPT cited us, on 4 October, cited our list of applicant tracking systems for small businesses, which averaged 56.8.

The lever test: how we scored each lever

The lever test takes each lever that LLM SEO guides recommend and asks one question of our own readings: did pages with the lever get cited more than pages without it? Each lever gets one of three verdicts: supported, contradicted, or not measured. A verdict covers our sites and our dates, not the lever everywhere.

Supported means pages with the lever drew more citations; contradicted means pages without it did as well or better; not measured means we have no pages without it, or too few readings to compare.

Every verdict below comes from these instruments. For Copilot, the AI Performance report of Bing Webmaster Tools, which also counts Bing’s AI summaries and some partner services and calls its figures a sample; we say “Copilot citations” for short. For rankings, Google Search Console. For Google’s AI Overviews, a paid data service’s database on wiserstaff.com and our own checks on riin-art.com. For ChatGPT and Google’s separate AI Mode, our own ten fixed questions each on 4 October.

Supported: testing the product yourself

Testing the product yourself is the lever with the largest gap in our readings, for Copilot. From 24 September to 3 October 2026, our five tested reviews drew 2,474 Copilot citations between them, 494.8 per review, while our 24 best-of lists drew 396, or 16.5 per list. No page was added in that window.

Two limits keep this verdict narrow:

• Testing was not enough on its own. Four of the five reviews were in Copilot’s list: Deel 1,363, Breezy 622, Gusto 286 and Tellent 203. The fifth, our Manatal review, has the same template, the same kind of screenshots and the same summary at the top, and has never appeared in that list.
• The citations came from searches about vendors. From 16 to 29 August, Bing tied 1,230 of 2,885 Copilot citations to a grounding query, the phrase its AI searched with; 1,211 of those, or 98.5%, came from phrases naming a vendor. We have no such split after 24 September. In our readings, a review was the page Copilot reached for when the search named a product. The lever may be “be a source on a product people search for”, with testing one possible reason the source is trusted; our readings cannot separate the two.

ChatGPT and Google’s AI Mode did not cite any review in our ten fixed questions each on 4 October, so this lever is a Copilot result.

Supported: dated prices in a table

Dated prices are the second lever our readings support, for Copilot, on one page. Our HR-software pricing guide, which dates every price row, drew 322 Copilot citations from 16 August to 3 October 2026, more than our ten other guides together (213), and is our only page a paid data service records as cited by Google’s AI Overviews.

The two answers we captured outside Microsoft are single readings, and neither shows our dated prices at work. The one ChatGPT answer that cited us, on 4 October, cited our list of applicant tracking systems for small businesses in an answer built on prices it presented as recent 2026 figures; we cannot tell where those prices came from. The one Google AI Mode answer that cited us cited our comparison of two recruiting tools, which sets a per-seat plan against a flat one.

The limit: we have one pricing guide, and no version of it without dates to compare against, so the gap may come from the topic as much as from the dates.

Contradicted: rank first, then get cited

“Rank well and AI will cite you” is the lever our readings contradict, for Copilot. None of our nine content pages with a top-10 average Google position in September 2026 made Copilot’s list of our most cited pages from 24 September to 3 October, and pages in that list averaged position 39.5 on Google, against 26.5 outside it.

The details, each from our own consoles:

• The best-ranked review never made the list. Our Manatal review averaged position 9.4 on Google in September, better than any other review, and has never appeared in Copilot’s list of our most cited pages.
• A page with no Google impression was cited. Our list of free applicant tracking systems had no Google impression at all in September, and Copilot cited it 6 times from 24 September to 3 October. Having no position, it is left out of the averages above.
• Bing search visibility did not come first either. Our Deel review, cited 3,879 times by Copilot from 16 August to 3 October, had no impression in Bing search from 15 August to 2 October, the last day in Bing’s search report when we read it.
• Crawling did not predict it. In late August, OpenAI’s bots read the site more than any other AI operator’s, and ChatGPT had not cited it once in our checks, as we showed in our crawl and citation readings.

Two cautions. Bing lists only our 20 most cited pages; from 24 September to 3 October they held 3,176 of the 3,328 citations, so a page outside the list may have been cited a few times, and “never in the list” does not mean never cited. And a page’s average position mixes every search it appeared for.

Structure and freshness: what did not separate cited pages

Summaries at the top, FAQ blocks, tables and schema did not separate the best-of lists in Copilot’s list from the others, and recent updates did not mark out our most cited reviews. All 42 of our content pages (5 reviews, 24 best-of lists, 11 guides, 2 comparisons) carry the same answer-first summary, so structure scores as not measured.

Counted on the source files of our 24 best-of lists as they stood on 9 October 2026, with the 11 in Copilot’s list from 24 September to 3 October set against the 13 outside it:

Tables, FAQ and summaries are close; the one large gap is length, about 700 words, which we did not test as a lever. On freshness: our Deel and Gusto reviews, last updated on 22 July 2026, remain our first and third most cited pages; a guide we refreshed on 16 September has never appeared in that list; and the Deel review went from 0 citations in Microsoft Clarity’s reading for 10 to 16 August to 552 in Bing Webmaster Tools’ reading for 16 to 20 August, without any edit.

This does not show that structure is useless. Every page has it, so our readings cannot tell what it adds; they can only tell it was not the difference. On freshness, pages left alone since July beat a page refreshed in September: a handful of pages, not a test.

Not measured: what our readings cannot tell you

Some of the most repeated LLM SEO advice is beyond what our readings can test. We have no controlled before-and-after measurement for structural changes, no brand-mention campaign to compare, and no group of pages without schema. And we can only partly explain the jump in our best-of lists’ Copilot citations, set out below.

The open questions, as they stand on 9 October 2026:

• Structural work, before and after. Our two readable cases, a table fix on 22 August and a pricing block added on 27 August, showed no visible gain in the days after, and neither can be isolated from other changes made in the same weeks.
• Mentions and links. The site has a handful of links from other sites, and from 25 to 31 August no citation came from a search containing its own name.
• An llms-full.txt file. Published in July, before our first citation reading.
• The jump in best-of citations. From 3 in 14 days, 16 to 29 August, to 396 in 10 days, 24 September to 3 October, while two lists were added, on 30 August and 2 September. The first of them drew more citations than any other list in the later window (97 of the 396), which explains part of the jump; our readings do not explain the rest.

A figure circulates that the overlap between Google’s top 10 and AI citations fell to 17 to 38%. It is not our measurement, and we have not traced it to a primary source, so we do not use it; our own count above measures something else, on Copilot.

What we would do on a new site

On a new site, we would test the products ourselves, publish prices with their dates in tables, and measure citations engine by engine. Our readings back those two levers for Copilot only; for Copilot, ranking first, structure and fresh dates did not decide which of our pages got cited.

• Observation: on wiserstaff.com, tested reviews and a dated pricing guide drew most Copilot citations, while our best-ranked pages, our structure and our fresh dates did not mark the cited pages out. On riin-art.com, in nine answers, Google’s AI Overviews followed Google’s ranking.
• Conclusion: LLM SEO is not one game. Copilot and Google’s AI Overviews rewarded different pages, and our readings do not show why; ChatGPT and AI Mode gave us one citation each, too few to call a pattern. A single LLM SEO score hides the engine that moved.
• Recommendation: name the engine before choosing a lever; for Copilot, put first-hand tests and dated prices first; for Google’s AI surfaces, our smaller Google readings suggest the usual ranking work still applies.

If you run one site and read your own consoles, you can run the same test on your pages for free, and you may not need the product this page links to. Our dated readings are on our figures page; what we automate to collect them is in our list of SEO tasks we run, cap or drop.