Core service

AI Recommendation Optimization

Improving how likely your business is to be named when a customer asks an AI assistant who they should hire — and measuring whether it worked.

Most marketing work is sold on the assumption that the activity is the outcome. Pages get published, schema gets added, directories get submitted, and everyone agrees it must be helping.

We start somewhere else: with a count. Before any work begins, we measure how often AI assistants currently name your business across the questions your buyers actually ask. Then we do the work. Then we run the identical measurement again.

That sequence is the whole product. It is also why we will decline to sell optimisation to a business that has not been measured first — there would be no way to tell you whether it worked.

Start with a baseline Read the methodology

What the work actually consists of

Four workstreams. Which ones get weight depends entirely on what the baseline found — a business that is invisible because its site cannot be parsed needs something different from one that is legible but has no third-party corroboration anywhere.

Make the business legible

Services, locations, credentials and service areas stated explicitly rather than implied. Structured data that matches the visible page. Crawler access for the systems you want to be read by. This is the floor — a business a model cannot parse cannot be recommended by one.

Make the facts agree

One name, one address, one set of services, everywhere a system might look. Conflicting information across your site, profiles and directories gives a model a reason to prefer a competitor it is more confident about.

Close the source gaps

Identify the third-party sources that repeatedly support the businesses AI already recommends in your category, and get properly represented on the ones that are genuinely relevant.

Publish something worth citing

Original expertise, real project evidence, specific local knowledge. A model asked for a recommendation has little reason to cite a page that restates what it already knows.

How an engagement runs

A repeating cycle, not a project with an end date. Each cycle re-runs the same 24 frozen prompts across 5 platforms, so the comparison is like-for-like.

1
Baseline
A prompt library built from your services, cities and buyer language, run and recorded in full. This is the number everything else is compared against.
2
Diagnose
Which prompts you lost, who won them, which sources the winning answers cited, and whether anything AI said about you was wrong.
3
Prioritise
A 30-day plan ordered by what the evidence suggests is constraining you — not by what is easiest to bill for.
4
Execute
Entity corrections, structural work on the site, source gap closure, and content built against the prompts you actually lost.
5
Re-measure
The identical benchmark, re-run. Movement is attributable because nothing about the test changed.
6
Report honestly
Current score, baseline, 90-day trend, platform differences, competitor movement, work completed — including work that produced nothing.

The most useful artefact: the source gap

Of everything an engagement produces, this is the piece clients act on first. It replaces “we should probably do more content” with a ranked list of named sources where your competitors are present and you are not.

Report structure — source types shown, not measured results

The columns of an AI Source Gap Report
Source	Competitors present	You	Relevance	Recommended action
Regional trade association directory	3 of 3	Absent	High	Join and complete the public listing
County-level local news	2 of 3	Absent	High	Earn coverage; cannot be purchased
Major review platform profile	3 of 3	Present	High	Present — correct services and service area
Industry certification register	2 of 3	Absent	Medium	Verify credentials are publicly listed
General-purpose business directory	3 of 3	Absent	Low	Deprioritise — low observed influence

What we will not claim

This category attracts overpromising, so it is worth being explicit about the boundaries before you buy anything.

We cannot guarantee that ChatGPT, Gemini, Claude, Perplexity or Google AI will recommend your business. Nobody controls the output of a model, and any agency offering that guarantee is selling something it cannot deliver.
We do not control ranking or placement in any AI system. We improve the information and authority signals those systems can draw on.
Completed technical work does not raise your Recommendation Score. It raises a separate diagnostic. If AI behaviour did not change, the score does not move and the report says so.
We cannot promise a timeline. Entity corrections can propagate quickly; authority accumulates slowly. You will see each cycle's result either way.
Some source gaps cannot be closed. Editorial coverage is not purchasable and some directories are not worth the effort. We will say which is which.

The individual services

Engagements are usually assembled from these. Baseline first, always.

Common questions

What is AI recommendation optimization?

AI recommendation optimization is the practice of improving how likely a business is to be named when an AI assistant is asked for a recommendation. It combines measurement of current recommendation behaviour, diagnosis of why competitors appear instead, correction of the business's machine-readable source of truth, and authority building across the third-party sources that influence answers.

What is generative engine optimization?

Generative engine optimization, sometimes shortened to GEO, refers to optimising for systems that generate answers rather than return links. Answer engine optimization and AI visibility optimization describe substantially the same work. Greenbug uses the term AI recommendation optimization because the commercially meaningful outcome for a local business is being recommended, not merely being cited.

How long does it take to see a change?

Honestly, it varies, and anyone quoting a fixed timeline is guessing. Entity corrections can be reflected quickly where the source updates quickly; authority signals accumulate over months. Because the benchmark is re-run on a fixed schedule, you will see whether anything moved each cycle rather than waiting for a verdict at the end.

Do I need to stop doing SEO?

No. AI systems rely heavily on crawlable, indexable web information, so search fundamentals remain part of the foundation. AI recommendation optimization adds work SEO does not typically cover: entity consistency across third-party sources, citation ecosystems, factual accuracy in generated answers, and measurement of recommendation outcomes rather than rankings.

What happens if the measurement shows no improvement?

It gets reported as no improvement. The Recommendation Score is computed from observed AI behaviour by a fixed formula with no manual adjustment, and completed technical work cannot raise it. A flat score is information — usually that the constraint is somewhere other than where the work went.

Start with the measurement

We will establish your baseline, show you who AI recommends instead, and tell you what the evidence says is constraining you.

Check my AI visibility Browse services