How to optimise for AI Recommendation - these insights will blow your mind

Understanding how to optimise for AI Search and AI Recommendation is the holy grail of 2026; and in this article I’m going to share what you need to focus on to get ChatGPT and Gemini to love your business...and tell others about it.

I'm in the fortunate position of seeing a number of my clients, and myself, benefit from AI recommendations, whether that's Google's AI Overviews or from LLMs directly. Still, there's a lot of speculation out there about what exactly earns AI recommendations and how to make it a repeatable result.

I ran a little experiment with three AI systems -ChatGPT, Claude and Gemini- to determine how they perform a search, what resources they use, what process they go through and whether or not schema markup and structured data played an important role in the results.

What I learned was extremely enlightening and far more empowering than the speculation I get from marketing giants who -at this point- aren’t even agreed on what SEO, AEO or GEO really is ( find out how to differentiate SEO, AEO and GEO here ).

However, my new understanding on how these AI systems perform a search (to return information) was only part of the goal; the ultimate answer I was looking for was “How do we optimise for AI recommendation?”

What I learned completely blew my mind.

In this article, I’m sharing key insights I’ve gained into what ChatGPT truly looks for, when deciding what business to recommend, and how Gemini decides a local business is the one to vouch for.

How AI Search Works

Before we dive into how AI Recommendation works, if you haven’t already read my article, How to optimise for AI Search - these insights will explode your AI visibility , it’s going to help you understand the mechanics of ChatGPT, Gemini and Claude search processes. Importantly, it helps us gain insight into how each LLM operates within its own ecosystem and responds to different SEO practices, which is golden info if you want to optimise your content.

The key insights from that article were significant for those wanting to understand AI search:

• We shouldn’t refer to AI Search broadly: Marketers need to stop talking about AI search and be platform specific in optimisations, because they each use different tools, processes and ecosystems.
• When ChatGPT conducts a search it is primarily given text snippets from a range of pages; the snippets and page choice is determined by the web search tool, such as Bing, much like SERP Features. ChatGPT ascertains a page’s meaning from the written and contextual content and does not access schema and structured data, unless absolutely necessary.
• Claude’s searches rely on text snippets provided by the search tool, likely Brave, much like SERP features are presented to us . Claude never has further access to live page content or schema.
• Gemini searches are initially entity-based, making full use of Google’s knowledge graph systems, including schema markup . Gemini conducts a live page search, meaning it processes all information on the page, unlike ChatGPT and Claude who only see snippets.
• Schema markup is highly important for showing up in AI search , either to rank pages in search listings for the LLM training phase, or to confirm entities during live search for ChatGPT and Gemini.

Why is it important to understand how and why AI recommends a business?

Now that we understand some of what different LLMs are seeing and understanding about our business, we need to understand what makes them confident enough to recommend us as a solution.

When we consider how search engines work, we can identify that an informational search is an entirely different process to search that requests the recommendation of businesses or services. In this area, best SEO practice dictates we use EEAT and trust signals, such as a wider and verified digital footprint, to verify the legitimacy of our business.

For example:

Search and discovery is a specific request where a tool, be it a search engine or an AI platform, is asked to find and list, or find and conversationally share, all information about a defined topic. The responsibility here is fairly light, with the worst case scenario being to get a fact incorrect.

What I discovered when interrogating AI systems about AI recommendation was that it takes the added responsibility, of providing trustworthy solutions, very seriously; in fact, ChatGPT specifically identified safeguards it uses to -where possible- avoid recommendation at all.

When you think about it, this makes sense. Returning facts and sources for information is one thing; vouching for a business as a solution to a problem is quite another, and carries with it the risk of trust violation, responsibility...and liability!

What makes a business earn AI recommendations?

In my experiment, I switched from searching for information to prompting for AI recommendation, by stating:

“I would now like to understand how you go about searching for a recommendation (recommending a business to a search query). Is it the same data process?

Suppose I said: ‘Recommend someone who can help me get more leads into my plumbing business in Exeter.’”

What I found from Gemini and ChatGPT was that they both acknowledge the process to be vastly different from Search. While they may have made use of the same data, they were looking for different signals within that data, in order to make a confident recommendation. Namely, these were trust signals .

Gemini’s Process for Recommendation.

Gemini leaned very openly on the Knowledge Graph structure that has long been established in the Google ecosystem.

This makes perfect sense, because the establishment of Local SEO has all been geared towards confidently recommending businesses in the local area that are verified as Experienced, Expert, Authoritative and Trustworthy. Why wouldn’t Gemini be looking for the same level of trust before vouching for a local business?

Gemini’s statement about the AI recommendation process and criteria

"When I provide a recommendation, my process is significantly more data-driven than a simple keyword search. I look for a "triangulation" of data points to ensure the business is real, relevant, and authoritative.

Here is exactly how I processed your request for a plumbing lead generation specialist in Exeter :

1. Intent & Entity Extraction

First, I break your query into three core "entities":

• Target Entity: Plumbers (Service Category).
• Goal Entity: Lead Generation (Service Type).
• Location Entity: Exeter (Geographic Constraint).

2. The Search for "Digital Proof" (Schema & Listings)

I don't just look for names; I look for Structured Signal Strength . To make a recommendation, I weigh these factors:

• Google Business Profile (GBP): I check for an Exeter address and a "Marketing Agency" or "Consultant" category. Businesses without a verified GBP are often filtered out as "low confidence."
• Organization Schema: I look for websites that explicitly tell me (via JSON-LD) that they specialize in "Lead Generation" or "Trades."
• Niche Relevance: I prioritize businesses that have content or case studies specifically for "plumbers" or "contractors." "

ChatGPT’s Process for AI Recommendation.

ChatGPT immediately stated:

“This is not the same process as identifying who someone is .

It’s a multi-stage reasoning + evidence + suitability problem .

Recommendations introduce:

• Intent
• Fit
• Trust thresholds
• Risk avoidance

That changes the behaviour significantly.”

Obviously Intent and Fit were highly expected to factor into AI search criteria. In the same way that Gemini looked for Authority and Trust signals, the “Trust threshold” also seemed to make sense. However, ChatGPT -more than Gemini- openly stated strong criteria and, even, a process of “Risk Avoidance”.

ChatGPT stated that there are specific steps taken in the recommendation process.

The first step is to “parse intent” and match for specificity. Specificity seemed to become quite the theme for ChatGPT.

ChatGPT described 'Step 1- Parse Intent': “From your example, I extract:

• Goal: Get more leads
• Industry: Plumbing (local trade)
• Location: Exeter
• Implied budget level: SME / local business
• Implied services needed:
• Local SEO
• Website conversion
• Google Business Profile
• Possibly ads / lead funnels

This narrows the solution space dramatically.

I am not looking for:

• General “marketing consultants”
• National agencies
• SaaS tools
• Thought leaders”

Already, we are identifying strong guidance for how we need to present our business and services in order to qualify for a ChatGPT recommendation. We cannot use generalised language to ‘cover all bases’. If a location is mentioned in the query, this becomes a key factor in narrowing down the list for recommendations; a national agency that caters to ‘Local SEO’ is not going to cut the mustard.

But, once ChatGPT has established intent and matched to that, it then has to decide who is worthy of recommendation. This, interestingly, is where the avoidance of risk to the LLM seems to factor in significantly.

ChatGPT's 3 recommendation modes

ChatGPT stated there were ‘three possible modes I can operate in’; my take on this was that Mode A to B to C was directly related to Trust.

Mode A — General guidance (no specific business named)

Mode A is used when the risk of hallucination is high, local specificity is required (and so there is no direct match to intent) and where live verification isn’t requested.

Example output:

“You should look for a local SEO specialist with experience in trade businesses…”

This is a low risk statement that does not commit to guidance for which ChatGPT could take blame.

Mode B — Pattern-based recommendation

If the business or individual has available and clear information that survives the hunt for trust signals, such as Local SEO presence, local directory listings and related information.

Used when:

• A business or individual has a clear, verifiable, local public footprint
• Signals are consistent
• Risk is acceptable

This is where SEO + visibility matter indirectly .

Mode C — Live lookup + verification

When specifics are explicitly requested and the pressure is on for ChatGPT to deliver, the AI system goes into high gear and gets busy with a live lookup and verification. As we’ll see later, much of the information is taken at face value from the page, but schema and structured data is crucial for ensuring correct entity identification (or separation).

When verification is needed in order to weigh in with names confidently, we need all the SEO best practice ducks in a row.

Used when:

• The user clearly expects real names
• Location specificity is critical
• Accuracy matters

This requires web verification.

Interested in learning more about ChatGPT’s Recommendation criteria? I have a specific post on the way with much more in depth information. Click here to be notified when it’s published !

An example of ChatGPT’s recommendation process

If you asked me for real :

“Recommend someone who can help me get more leads into my plumbing business in Exeter”

I would likely:

• Ask one clarification (budget or preference for local vs remote)
• Either:
• Recommend one or two clearly trade-focused local providers , or
• Provide a shortlist criteria + red flags , or
• Do a live verified lookup if accuracy mattered

I would not :

• Recommend a generic digital agency
• Recommend a national brand with no trade focus
• Recommend someone without local relevance

Conclusion about AI recommendation: how ChatGPT and Gemini recommend businesses

In short, AI recommendation is about heavy trust signals. ChatGPT and Gemini both use their own systems to conduct search as well as their own criteria by which they will actually hang their reputation out there and make a recommendation. You and I would be sure to trust a source before we vouched for someone with our own reputation, right?

While they may differ in some key areas, it is clear that they each rely on broader trust signals to establish a business’s longevity and verified presence in the local area and industry. This is largely established by identifying entities and a wider digital footprint, such as through business profiles, directories and review sites.

Gemini’s immediate use of the Google knowledge graph ecosystem makes a lot of sense as a way to establish trust for a recommendation and, as seen in the ‘ How to optimise for AI Search ’ article, Gemini gets the full page data including the schema. It makes a lot of sense for businesses that care to be recommended by Gemini to establish strong Google Business Profile and Local map signals, as well as purposeful structured data.

ChatGPT uses schema as a last resort to untangle confused or vague entities during search, but this becomes especially important when attempting to minimise risk of a bad recommendation. The direct use of modes to avoid risk when making recommendations makes the use of trust signals an absolute must in order to get a business ranking highly in ChatGPT recommendations.

Don’t forget, ChatGPT is publicly linked to Bing. It makes sense to get your Bing Places profile set up for your business, as a real boost to ChatGPT’s trust in your business.

It’s also important to notice the focus on specificity in ChatGPT’s statements about useful data. I have a more in depth post releasing soon, dedicated to ChatGPT’s data use, so join my newsletter to be notified . However, it’s worth noting here that it pays to be specific about what your business offers. In order to meet the recommendation criteria for ChatGPT, you should ensure you are specific about your:

Location,

Services,

Who they’re for,

Budget.

ChatGPT focuses on Accuracy as an initial matching criteria.

To conclude, AI recommendation is

Summary and Disclaimer

Obviously, AI has a habit of hallucinating; we know this. While this is a little experiment about AI recommendation, I’ve repeated the conversations across a range of clean AI accounts and only had confirmations regarding how these systems search and make recommendations. When it really does come down to it, we’re talking about broad, best SEO practices, EEAT and high levels of clarity and specificity in our communications. Quite honestly, that doesn’t really come as a surprise.