Readable Versus Recommended: What AI Search Really Rewards

🇹🇷 Türkçe: Bu yazının Türkçesini oku →

If you run a brand of any size, there is a quiet experiment you can do this afternoon. Open ChatGPT, Perplexity, Claude and Google’s AI Mode in four tabs, type the same category question into each, and read what comes back. A credit card for small businesses in Turkey, a mid-market payroll system, a courier for same-day in Istanbul.

Four years ago that exercise would have told you something about keywords. Today it tells you whether your brand is in the room at all, whether anyone is being told to prefer you, and whether the model agrees with the version of your company you have been paying to put online.

A lot of the money moving around AI search right now is spent making sites readable to crawlers. That work matters, we do it, and we charge for it. But it is not the thing that gets you cited, and a lot of boards are about to discover they commissioned the first part of the job and were not told there was a second part.

So I went and asked five people who have published serious work on this question what the second part actually is, and where the money for it comes from.

The five voices

• Jodi Cerretani, CMO, WordPress VIP.
• Paxton Gray, CEO, 97th Floor.
• Mike King, Founder and CEO, iPullRank. Named Search Engine Land’s 2025 AI Search Marketer of the Year.
• Rand Fishkin, co-founder, SparkToro. Previously co-founder of Moz.
• Aleyda Solís, SEO and AI search consultant, founder of Orainti.

Every quote below is from a published, linkable source. Nothing here is paraphrase presented as speech, and nothing came from a private conversation. The links go to the original material.

Being readable is a floor the industry just finished pouring

Start with the line that frames the whole argument. Cerretani, speaking in the WordPress VIP and 97th Floor enterprise AI discoverability webinar on 22 September, put it in a single sentence that most of the industry has been trying to articulate for a year:

Being readable isn’t the same thing as being recommended.

Jodi Cerretani, CMO, WordPress VIP, session 2, 8:43

She goes on to use a library analogy that is worth stealing: information architecture is the floor plan, structured content is the label on the individual book, governance is the librarian who keeps the labels accurate. Pour the floor, label the books, employ the librarian, and a visitor can find what they came for. None of that, by itself, tells the visitor which book to borrow.

Mike King writes the engineering half of the same argument. From the article announcing his Search Engine Land 2025 AI Search Marketer of the Year award, published on iPullRank on 27 October:

SEO was never meant to be static. It was never about chasing 10 blue links. It was always about decoding how meaning moves through machines.

Mike King, Founder and CEO, iPullRank

And in a long essay for Advanced Web Ranking on what he calls relevance engineering, King draws the role the industry is actually moving into:

we’re not just mechanics tweaking engines, we’re engineers building the actual systems.

Mike King

Read the two together and the picture is consistent. Readable is the engineering achievement that lets a model understand you at all: one H1, clean hierarchy, schema baked into the template, a sitemap that matches what you are actually promoting, a server layer that does not reject Claude or GPTBot by accident. Cerretani’s own disclosure that wpvip.com’s server layer was blocking one of Anthropic’s crawlers while the team was ranking themselves on technical readiness is the honest version of how often that foundation is quietly broken even at the top end.

What a model recommends is largely what the web already says about you

This is where Fishkin and Solís push the argument past the on-page work the industry mostly knows how to sell.

Fishkin’s SparkToro research team, with 600 volunteers, ran roughly 3,000 prompts through ChatGPT, Claude and Google’s AI tools and compared the brand lists each one returned. The headline is uncomfortable: fewer than one in a hundred runs produced the same list of brands, and fewer than one in a thousand produced that list in the same order. His read, as summarised in his Advanced Web Ranking interview, is that the only companies showing up consistently are the ones that earned a presence on the open web first:

Brand is different. Brand is creating memorable campaigns. It’s standing out in people’s minds.

Rand Fishkin, co-founder, SparkToro

Solís has done the quantitative version of the same claim. Her analysis of 15 leading brands across Google AI Mode, Gemini and ChatGPT, using Semrush Enterprise data, reports that owned domains are 17.7% of the top cited sources in SaaS, 30.4% in ecommerce and 20.6% in finance. The rest, between two thirds and four fifths of the citations a model draws from when it mentions you, sit on properties your brand does not own. Her summary line is as blunt as it should be:

AI search is a third-party citation problem with an on-page corroboration base.

Aleyda Solís, founder, Orainti

Fishkin’s inconsistency finding and Solís’s citation share reach the same operational conclusion from two angles. If your plan to be recommended is to ship a tidier sitemap, you are optimising a floor the recommender cannot see. Reddit threads, YouTube reviewers, industry publishers, partner marketplaces, Quora answers your lawyers would rather you did not write: that is the material a model averages when it decides what to say about you.

Paxton Gray, in session 1, had a client case that lands this at board level. An unnamed venue management software vendor wanted to move down-market and asked an assistant to describe them. The answer, at 4:08: the IBM of venue management software, very well established, but clunky, hard to use, not for small venues. The site itself said nothing of the sort. The web around it did, and the model averaged.

Prompts are not keywords, and audience research is what replaces them

Gray’s own contribution to the argument is a correction to how the industry is costing the work. Keyword lists shaped a decade of content budgets and told writers what to produce. The analogue in AI search is not a longer keyword list.

the big mistake a lot of people in the industry are making is they are taking prompts and they are equating those with keywords, but they are not the same thing.

Paxton Gray, CEO, 97th Floor, session 2, 9:55

A keyword is a short string. A prompt is a sentence of context about a situation, a budget, a deadline, and the question underneath. The reader is not asking payroll software, they are asking which payroll system fits a Turkish company of 180 people, bilingual, with offices in two cities. If your content was built around the first query, the model does not have the pieces it needs to answer the second one with you in it.

Gray’s recommended replacement is the fundamentals the industry forgot:

gone are the days of the traditional keyword research as ways to establish and build your strategy around AI, and now we need to go back to fundamentals, which is audience research.

Paxton Gray, session 2, 11:17

The practical consequence for a content budget is that the unit of planning gets bigger. Instead of a hundred articles chasing a hundred phrases, you need the twenty or thirty questions a real buyer in your category actually asks, modelled once and told from several angles, with the proof points attached in a form a machine can lift.

Governance decides whether readability compounds or decays

One further contribution before the build. Cerretani returns to the governance side of her library analogy with a line most organisations would prefer not to hear:

do you have a website software that can allow you to do this automatically, or are you dependent on your human resources to make sure to do all these things? If the answer is the latter and not the former, you might be in trouble.

Jodi Cerretani, session 2, 6:50

This is the sentence that quietly settles the platform question. If keeping content machine-readable depends on an editor remembering to add schema in a WYSIWYG, the structure is dead within six months. If a template emits the right markup whether or not the editor thought about it, the structure holds while your team changes and your agency contract turns over. The second one is what Cerretani means by software doing the job, and it is also where WordPress sits as a mature platform: blocks, templates, pattern libraries and must-use plugins are the surface that enforces it.

Her summary line is the right place to end this section:

if you have infrastructure and it’s all really buttoned up but no content strategy, that means your brand is readable but it’s not necessarily going to be cited or recommended. On the flip side, if you’ve got killer content strategy but no infrastructure, it’s a great story that no machine can read.

Jodi Cerretani, session 2, 24:04

What this looks like as a build inside a WordPress and WooCommerce estate

Five voices, one argument, and the usual gap at the bottom of a strategy post: what any of this costs to actually ship. Here is the version I would scope for a mid-market brand with a few hundred SKUs and a public catalogue the organisation genuinely cares about.

One, model the content once and let the template emit the schema. Replace free-text description fields with native block or ACF fields for the handful of attributes a model needs: who it is for, what problem it solves, what it costs, what it integrates with, what the alternatives are. The template renders JSON-LD at publish time. For a 500 to 2,000 SKU WooCommerce catalogue this is two to three weeks of data modelling and template work, plus a weekend of migration. Nobody writes schema into a post again.

Two, audit the sitemap against what you are actually promoting. A sitemap built in 2022 lists what you shipped in 2022. Prune the pages that teach a model the wrong thing about you using Gray’s method: pull visits and engagement per page, run a sentiment pass to find where a model is stating something off about your brand and citing your own site for it, start there. First pass across a 500 to 2,000 URL estate is three to five days with analytics. The slow pass is a quarter’s work.

Three, repair the crawler layer. Pull a week of server logs, filter for the user agents of GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot and Google-Extended, compare the 200 count to the 403 or 429 count. On almost every estate we have looked at, something in the WAF, the compliance plugin or the CDN rules is throttling at least one of them. A day to find, half a day to fix, a day to confirm.

Four, invest where Solís tells you to. If two thirds to four fifths of citations are on properties you do not own, treat the external surface as a budget line, not an afterthought. Industry publishers, partner marketplaces, vertical review sites, a handful of answerable Quora and Reddit threads where your category is actually discussed, a steady drip of attributed quotes from your executives in trade press. This is the line most mid-market budgets do not have.

Five, decide now who holds the role. Marketing, content, SEO or web can all hold it. What is not survivable is nobody holding it. Pick one, write it down, give them access to analytics and the publishing stack, and set a retest cadence. Ninety days is the right rhythm for the first year.

End-to-end: six to eight weeks for the first pass, then two to three days a quarter. Half engineering, half editorial. The budget most organisations miss is the external one.

The honest part: measurement still lags the whole thing

Nobody on this panel is pretending the measurement is solved. Gray is on the record that it is the industry’s biggest open problem, that most current reporting is correlation and not causation, and that his honest answer is a billboard analogy: ask whether branded search is up, direct traffic is up, conversion improved. Cerretani is willing to say that current measurement tools, hers included, run on JavaScript and therefore see humans, not AI crawlers. We will come back to measurement on its own after WordPress VIP’s third webinar on 20 October.

In the meantime: if you have been paying for an AI-readable site, you have bought the floor. It was real work, it needed doing, and the invoice was fair. What you have not bought, and what nobody told you was a separate purchase, is the thing that gets you cited rather than just understood. If you want to talk about where that budget should live and what we would actually build, that is a conversation we have most weeks, and I would rather have it with you before the next quarterly review than after it.