Ask an assistant the same question about the same Miami business in English and then in Spanish, and you will often get two different answers assembled from two different sets of sources. Almost nobody checks both, which is how a firm ends up well cited in one language and effectively absent in the other.
Search did not stop working the way it always did. Ten blue links are still there, still measurable, still where most of the traffic comes from. What changed is that a layer settled on top: a generated answer that reads the sources for the user and hands back a paragraph, naming three or four of them along the way.
That layer needs its own vocabulary. Ranking is position in a list. Visibility in a generated answer is whether your page was one of the handful the model drew on and named. The two correlate loosely and fail independently.
A list you scan against an answer you are handed
On a classic results page the engine retrieves and you judge. Ten candidates appear, you read the titles, you pick; the metric that matters is where your page sat in that stack.
In a generated answer the judging happens first. The system retrieves documents, reads them, writes a synthesis, and names a few as sources. If your page was not retrieved, or did not survive the synthesis, you are not in fourth place — you are simply not there.
| Aspect | Classic result list | Generated answer |
|---|---|---|
| What the user sees | Ten candidates to choose from | One paragraph plus a few named sources |
| Unit of success | Position | Inclusion |
| Failure mode | Ranked too low to be seen | Not drawn on at all |
| Measurement | Impressions and clicks, reported | No first-party record |
| Stability | Same query, similar list | Same query, sometimes a different answer |
That last row is not a detail. Retrieval shifts with phrasing, with who is asking, and with whatever was indexed that week, so two people typing the same question can get different sources on the same afternoon.
Cited is not ranked, and ranked is not cited
A page at position three has proven that a ranking system judged it a strong match. A page named inside a generated answer has proven something else: that a model retrieved it, read it, found a passage that answered the question cleanly, and considered it worth naming.
Plenty of pages pass the first test and fail the second — a service page that ranks but is mostly assertion, with no passage an assistant can lift. Pages also pass the second while failing the first: an explanatory page that never cracked the top ten but keeps turning up as a source.
So you cannot infer one from the other. A firm that spent three years on rankings may already be cited and never know it. Guessing is worthless here; checking is manual work.
Where the answer material comes from
Answer engines do not invent claims about a Coral Gables dental group or a Doral freight forwarder; they retrieve documents and synthesize. Which documents get retrieved is the whole game, and the pattern is consistent enough to plan around.
Review and listing platforms
Structured, repetitive, machine-readable statements: category, area served, hours, sentiment. Easy to draw on.
- Heavily weighted in English
- Far thinner in Spanish
Trade press and local publications
A mention in an industry title outweighs a directory entry: written prose about you rather than a form field.
- Rare but durable
- Often the deciding citation
Your own explanatory pages
The one source you control. Pages that state facts plainly — what the service is, who it is for, what it costs, where it is delivered — get lifted. Pages built on adjectives do not.
- Directly editable
- Language-specific by construction
Forums, threads and Q&A
Real people comparing providers. Unpredictable, unbuyable, and in Spanish-language Miami queries often the largest share of what exists to retrieve.
- Impossible to control
- Worth reading regardless
Three of those four are somebody else's property, and the one lever you hold has to be pulled in each language separately.
Ask it in English, then ask it in Spanish
This test costs nothing and almost no firm here has run it. Take a question a customer would actually ask — the best pediatric dentist in Coral Gables, a Brickell immigration attorney handling investor visas, a Doral customs broker for Caribbean shipments — put it to an assistant twice, once in each language, and read the answers side by side.
They will usually differ, and not in the trivial way of one being a translation of the other. They name different businesses, and they are built from different sources. The English answer tends to come from large national review platforms and English-language publications: a deep, well-structured, heavily duplicated pool. The Spanish answer comes from a thinner, more local set — a few Spanish listings, community discussion, whatever regional coverage exists, and the Spanish pages of the businesses themselves, where they exist at all.
| Dimension | English-language answer | Spanish-language answer |
|---|---|---|
| Source pool | Deep, national, well structured | Thin, local, uneven |
| Typical citations | Review platforms, trade press | Community threads, the firm's own pages |
| Weight of your own site | One voice among many | Frequently decisive |
| Competitors named | The nationally visible ones | Whoever wrote in Spanish at all |
| Who is checking it | Some firms | Almost nobody |
Read that table the right way round. A thin Spanish source pool is not bad news for a Miami business — it is the most exploitable asymmetry in this market. When the material available to answer a Spanish-language question is scarce, a firm publishing good Spanish pages is often one of the few things there is to cite.
- Run the pair test on ten real questions. Not brand questions — the questions a customer asks before they know your name. Record which businesses and which sources each language produces.
- Check whether your own site appears at all in Spanish. Machine-translated stubs do not get drawn on, so the one lever you control has been handed to nobody.
- Note who shows up in Spanish and not in English. Those are your real competitors in the half of the market most agencies never look at, and rarely the firms you track on rankings.
- Repeat it monthly on the same questions, with the exact phrasing recorded. Answers move, and changing a few words changes the retrieval. Without a fixed list you are comparing two different tests and calling it a trend.
When the buyer asking is in Bogotá
A business selling across borders has a second version of this problem, harder to see because the question is not asked in Miami. A buyer in Bogotá, Caracas, San Juan or São Paulo researching a South Florida supplier gets an answer built from that buyer's context — regional directories, Latin American trade discussion, chambers of commerce, local coverage. Sources the Miami firm has never read.
That matters for the industries this city runs on. Freight forwarding, clinics treating patients who fly in, immigration practices, real estate serving overseas buyers, marine services — all get researched from outside the country before anyone picks up the phone, and none of it shows up in a Miami-shaped view of the market.
You cannot observe an answer served to somebody in Colombia, but you can observe the demand underneath it. The country breakdown in the Search Console views is the nearest honest proxy: which countries produce impressions and clicks, on which pages, for which queries. A clinic that finds a fifth of its impressions coming from outside the United States has learned something concrete about whose questions an assistant is answering with — or without — its material.
Read that split as demand, not noise. Its fastest use is negative: a country with meaningful impressions and almost no clicks means the landing page is usually in the wrong language.
What a visibility estimate actually is
Since no engine reports citations, any product offering a visibility figure infers one. Understanding the method tells you how far the number can be pushed before it breaks.
Six views built on inference, not on a feed
For teams that want a structured read on the generative layer and know what kind of number they are reading.
- A competitiveness score with a market circle. Your domain placed against top-tier, mid-tier and niche competitors, rather than a bare number with no reference points.
- Model-generated market context. Positioning, a traffic estimate and stated opportunities — a written read on the market, produced from what the model can see.
- Query research and intent classification. Which questions exist around your subject and what the asker wants. Input to the pair test, not a replacement for it.
- Pages flagged as levers. URLs marked as worth expanding or worth linking to internally, which turns an abstract score into a work list.
- Competitor strengths and content gaps. Where rivals are covered and you are not. The gaps are usually where the citations went.
- A global visibility value across the AI search landscape. One aggregate figure, useful as a direction of travel and only that.
The construction, in outline: assemble representative questions for a market, put them to models, record which domains appear and how prominently, aggregate, express the result as a score with competitors around it. Every step is defensible. None of it is measurement in the sense that a click is measurement.
Within those limits it does real work. Comparison against three named competitors beats the absolute figure, and direction over three months beats any single reading. The content-gap and page-lever views point at pages rather than at a number, which is useful whether or not you trust the score.
Writing pages an answer engine can use
None of this is a trick, and none of it conflicts with writing for people. Pages that get drawn on share a shape: they state things, in one place, without making the reader infer. What a model lifts is what a hurried reader wanted anyway.
Say the fact, then explain it
Lead each section with the claim in a plain sentence. A passage that answers in isolation can be quoted; one that builds to a conclusion over five paragraphs cannot.
- One idea per paragraph
- No withheld conclusions
Numbers, places, conditions
Prices, service areas, turnaround times, what is included. Specifics survive synthesis; superlatives are discarded because every competitor supplies them too.
- Name the areas served
- State the real constraints
Two languages, two real texts
A machine-translated Spanish page adds no new material to a thin pool. Written properly, it is often the strongest source available in that language.
- Write, do not translate
- Use the vocabulary the trade uses
Question-shaped headings
Headings phrased as the question a customer asks, with the answer immediately after. Retrieval handles that format best and readers scan it fastest.
- Real questions, not slogans
- Answer in the first sentence
Off the page, the work is what it always was: matching business details across listings in both languages, coverage in the publications your market reads, and links from sites with genuine traffic. The placement network behind the campaign side spans more than 230,000 websites, and the same links that support ranking raise the chance your pages are retrieved at all.
| Change | Effort | Helps ranking | Helps citation |
|---|---|---|---|
| Question-shaped headings with direct answers | Low | Somewhat | Yes |
| Specific figures, areas and conditions on service pages | Low | Somewhat | Yes |
| Spanish pages written rather than translated | High | Yes | Strongly |
| Keyword density work on existing copy | Low | Marginal | No |
The last row is deliberate. Habits surviving from an older era of optimization do nothing for inclusion here, and that time is better spent rewriting one service page in real Spanish. The on-site suggestions produced by the panel point at the same fixes, and in FullSEO they can be routed through a human review mode before publication.
Questions that come up
Can I find out how often an assistant cites my site?
Not exactly, and not from any first-party report. Fix a list of questions, ask them on a schedule, record which sources appear, and watch the pattern. Any tool offering a citation count is doing a version of this.
Should we shift budget from ranking work to this?
No. Clicks from the result list remain the majority of measurable traffic and the only part with reported data behind it. This layer sits on the same foundations — indexable pages, real content, credible links — so work that improves one usually improves the other.
Our Spanish site is a translation of the English one. Is that a problem here?
A missed opportunity more than a fault. A translated page rarely adds material that was not already in the pool, and where the pool is thin an original Spanish page is disproportionately likely to be drawn on.
How long before any of this shows movement?
Same horizon as the rest of the work: first measurable movement typically appears four to eight weeks in, and this layer is slower and noisier because the sample itself fluctuates. One month of readings is not a trend.
Putting it in the report without overclaiming
Reporting is where the damage gets done, because a score that was honest in a spreadsheet becomes a promise the moment it lands in a deck. The fix is not to hide the number but to label it, place it after the measured data, and describe it as an estimate.
Where the estimate goes in the report
A layout that keeps recorded data and inferred data visibly apart.
- Reported figures first. Clicks, impressions, average position, keyword movement in and out of the top three, ten and thirty, and the country and device splits.
- The estimate in its own section. Visually separated, never interleaved with measured data, and never on the summary slide as though it were a KPI.
- Direction over three months. One reading is a sample; three consecutive readings against the same question list are an argument.
- The question list attached. Forty fixed questions in both languages, with what each answer named. That appendix is what makes the section auditable.
Underneath the reporting, the ordinary work continues. This layer rewards what was always worth doing — findable pages, real content, credible mentions — with one addition specific to this city: doing it in both languages is the difference between being one of many possible sources in English and one of the few available in Spanish. More walkthroughs sit on our blog, and the audit above is part of what we run for clients.
On the campaign side there are two levels. AutoSEO at $149 per month per domain handles keyword discovery, link building and on-site suggestions automatically. FullSEO at $500 per month per domain adds manual keyword selection with automatic fallback, placement against a target domain rating, and the human review mode — worth it on whichever version carries the revenue. The Stream assistant answers against the project's real data rather than in general terms.
Start with the pair test: twenty questions your customers really ask, once in English and once in Spanish, written down with the sources each answer named. Then connect the domain and open the AI Analytics views to see where the estimate places you against the competitors it names. The gap between what the two languages return is usually the most useful thing on the page, and for most firms here it is the first time anyone has looked.