What AI Assistants Actually Cite: A Teardown of the Sources Behind Local Business Recommendations

Everybody Is Asking How to Rank. Almost Nobody Has Read the Answer.

A large amount of advice is now being sold on the premise that you can optimize for AI assistants. Very little of it starts where it should, which is with the observable output: when you ask an assistant to recommend a business in a category and a place, it names some businesses and, frequently, it shows you what it read. Those citations are not a mystery to be theorised about. They are a list, sitting on the screen, that anyone can collect.

This piece does two things. It sets out a structure for what gets cited — five tiers we call the citation ladder — and it gives you the protocol to run the teardown on your own category, so the finding is reproducible rather than something you have to take on faith. It is the source-level companion to our earlier work on demand that analytics cannot see. That piece was about not being able to measure a channel. This one is about the machinery inside it.

The Citation Ladder: Five Kinds of Source, in the Order They Get Used

These are tiers rather than a ranking table. The useful property is not their exact order on any given day; it is that they behave differently, respond to different work, and give a business very different amounts of control.

Tier 1 — Structured business records

The map and profile layer: the major business profile services and their equivalents on other platforms. These are the most machine-readable documents about your business in existence — category, address, hours, service area, attributes, review count, all in fixed fields. They are consulted constantly and they are almost fully under your control, which makes them the cheapest work on this list. They are also the least differentiating, because every serious competitor has one too. Completeness here is table stakes rather than an advantage.

Tier 2 — Category-coverage documents

Directories, industry-body member lists, licensing and certification registries, marketplace category pages. This tier is the workhorse, and it is the one most small businesses underestimate, because from a human point of view these pages are unappealing and the information on them is often out of date. That is beside the point. They describe the entire category in one place, which is a structural advantage explained in the next section.

Tier 3 — Editorial and local press

City publications, local news, trade press, genuine roundup articles. This is the highest-leverage tier and the one small businesses almost never pursue, because it cannot be bought or submitted to — it has to be earned by being interesting to a person with an audience. A single credible local article naming your business does more for how you are described than a year of posting on your own domain, and it keeps working long after it is published.

Tier 4 — Community discussion

Forums, community threads, question-and-answer posts, neighbourhood groups. Heavily weighted, because it reads as unpaid and specific. Also almost entirely outside your control, and the one tier where intervention tends to backfire: communities are good at detecting a business talking about itself, and the reputational cost of being caught exceeds any plausible benefit. Read this tier. Do not farm it.

Tier 5 — Your own website

Cited least often in the recommendation itself, and cited very reliably once your name is already in play. It is where the assistant goes to establish what you actually do, whether you serve the area, what your hours are, and whether the description assembled from tiers one through four is correct.

The Finding That Reorders Everything: You Are the Confirmation, Not the Recommendation

Read the ladder from the top and a pattern falls out that contradicts most of what is currently being sold.

The shortlist is assembled almost entirely from documents you did not write. By the time your own website is consulted, the question has already narrowed from "who should I recommend" to "is this one what it appears to be". Your pages are doing verification work — checking the category, the service area, the specialism, the hours — on a candidate that other people’s documents put forward.

That has an uncomfortable implication for the standard advice. Publishing more content on your own domain improves the confirmation step, which is real and worth doing, and does very little for the shortlist step, which is the one that decides whether you are named at all. A business can write forty excellent pages and remain invisible to assistants in its own city, because the work it did addressed the part of the process it was never failing.

The corollary is the practical instruction of this entire article: to get on the shortlist you have to be described by other people in documents that cover your whole category. To stay on it once named, your own site has to confirm what those documents say. Those are two different jobs, and most businesses are only doing the second — a point developed from the business owner’s side in why ChatGPT does not recommend your business.

The Coverage-Document Bias: Why Directories Beat You With Worse Information

The most common objection to the ladder is that directory data about a business is frequently stale, thin and worse than what the business publishes itself. That objection is correct and it does not change the outcome, for a reason worth stating precisely.

When a system has to produce several candidates in a category, a single document naming twelve of them is a more efficient source than twelve documents naming one each. One retrieval returns the whole answer, already grouped by the category and the place, in a consistent format. This is a property of how the retrieval step works rather than an editorial judgment about quality, which is why it survives being pointed out.

Two things follow. First, a mediocre listing inside a comprehensive document frequently outperforms an excellent page that stands alone. Second, and more useful, the highest-return work available to most small businesses is not publishing something new — it is getting accurately described inside documents that already cover their category. Association member lists, licensing and certification registries, legitimate trade directories, supplier and partner listings, local chamber pages, and the roundup articles a local publication updates every year. Most cost nothing to appear in, most take under an hour, and most businesses have never audited which ones they are missing from.

Eligibility Before Ranking: The Three Facts That Decide Whether You Can Be Named

There is a step before any of this that most optimization advice skips entirely, and it explains the majority of cases where a business is completely absent rather than merely losing.

To be named in a recommendation, three facts about your business have to be establishable: an unambiguous category, a clearly stated service area, and at least one attribute that distinguishes you from the other candidates. Category answers whether you belong in the question. Service area answers whether you belong in the place. The distinguishing attribute answers why you rather than the other nine.

The third one is where most businesses fail, and they fail by trying to be eligible for everything. A page that says you serve all industries and offer every service supplies no attribute at all, which makes you a candidate that cannot be argued for. Specificity is not a marketing preference here. It is the input that makes a recommendation constructible.

This is also the practical bridge to the other half of the work. A business absent from every category-level document is not ranked low; it is not a candidate. That is a different problem with a different fix, and it is the one that shapes how we scope answer engine optimization engagements — eligibility first, then description, then the site that confirms both.

The Consistency Penalty Nobody Names

Here is the failure mode we see most often in businesses that are doing everything else right, and it costs nothing to fix.

Your business profile lists one primary category. Two directories list a different one. Your website describes you a third way. Your service area is three counties on one record, a city on another, and unstated on your own site. None of these are errors anyone would notice, and a human reader would reconcile them without effort.

A system assembling an answer does not reconcile them. It has conflicting evidence about a candidate, and when there are eleven other candidates whose records agree with each other, the cheapest resolution is to name one of those instead. You are not penalised for being wrong. You are passed over for being ambiguous, which is worse, because nothing in your analytics will ever show it happening.

The fix is an afternoon of clerical work most businesses never do: pick one category label, one service-area description and one business description, then make every record on the internet say the same thing. It is the least interesting recommendation in this article and reliably one of the two or three most effective.

Run the Teardown Yourself: The Monthly Protocol

The point of a framework is that you can test it against your own category rather than trusting ours. This takes about forty minutes the first time and twenty thereafter.

Build a fixed prompt set of ten

Write ten questions a real customer would ask, and freeze them — the value of the exercise comes from asking identical questions every month. Cover four shapes: the plain category-and-place request, the qualified version with a constraint that matches your specialism, the problem-first version where the customer describes a symptom rather than naming a service, and the comparison version asking who is best for a specific situation. Include one question your ideal customer would ask and one your worst-fit customer would ask, because being named in the second is a signal your category label is wrong.

Ask three assistants, logged out

Use three different assistants and sign out of all of them, or use a private window. Logged in, you are being answered partly on the basis of your own history, which will show you your own business more often and quietly ruin the exercise. Run all ten prompts on each, in one sitting.

Log four columns, then count

For every answer, record the prompt, the businesses named in order, every source domain cited, and which of the five tiers each source belongs to. Then produce two counts: how many citations fell into each tier, and which specific domains appeared more than twice across the thirty answers. That second list is the important output. It is the set of documents that describe your category to the machines your customers are asking, and for most businesses it is between four and ten domains — a target list short enough to act on this quarter.

Reading Your Own Distribution

Three patterns come up repeatedly, and each points at a different piece of work.

If your category is dominated by tier two, the work is placement: get accurately listed in the specific directories and registries that keep appearing, and make sure what they say about you matches everything else. If it is dominated by tier three, the work is being worth writing about, which is slower, harder and produces a far more durable position than any listing. If tier four is heavy, your category is decided by word of mouth in public, and the honest move is to improve the thing being discussed rather than to attempt to influence the discussion.

And if your own domain appears in the citations at all, that is a good sign about your website’s legibility — it means your pages are readable enough to be used as evidence. Whether they are also structured to convert the visitor who arrives already half-decided is a separate question, and one worth checking, since assistant-referred visitors behave differently from search visitors in ways covered in traffic without leads.

What This Method Cannot Tell You

Stating the limits is not a disclaimer here. It is the part that separates a method from a pitch.

Answers are not deterministic. The same prompt, asked twice in a day, can return different businesses and different sources, and the wording of a question changes the answer more than most people expect. Results vary by region and by device. Assistants differ in whether they search live or answer from what they have already absorbed, so a change you make will show up on different timelines across the three. Citations shown are not a complete account of what informed the answer, only of what was surfaced. And no vendor, ourselves included, can promise you a placement, because there is no placement to sell.

What the protocol does deliver is a trend line and a target list, which is more than most businesses have. Run it three months in a row and you will know whether you are being named more often, which documents decide your category, and whether the work you paid for touched any of it. That is a considerably lower promise than the industry is currently making, and it has the advantage of being true.

Where the Money Should Go if the Ladder Is Right

The budget instruction that follows from all of this is unglamorous and cheaper than what is usually proposed.

First, spend nothing and fix the consistency problem: one category, one service area, one description, everywhere. Second, spend a small amount of time on placement in the four to ten documents your teardown identified. Third, spend real effort on being worth writing about locally, which is the only tier that compounds. And fourth — not first — make your own site the confirmation it needs to be: pages that state plainly what you do, where, and for whom, in text a machine can read rather than in images and slideshows.

That fourth item is the part we build. A site that reads cleanly to an assistant is the same site that reads cleanly to a customer arriving already half-convinced, which is why the build, the search foundation underneath it and the intake layer that catches the enquiry are one job rather than three. Done for you, built with Claude Code — and measured against a teardown you can run yourself, which is the part we would rather you kept doing.