AI search
Google likes you. ChatGPT has never heard of you.
The gap between classic rankings and AI answers worries many site owners. Here is what is known, what is guessed and what a link API can help with.
SEOCheck research team ·

Honest status
Nobody has a proven formula yet
Anyone selling a guaranteed recipe for ai search visibility is guessing. The reasonable bet is that being clearly known across the web helps in both worlds, and links are one visible trace of that.
- Backlinks are not dead. Their role is unproven either way.
- We do not track AI answers. No product here does.
- Treat GEO as extra work on solid SEO, not a replacement.
What is different about AI answers
A classic result list shows ten links and lets you choose. Simple. An AI answer writes a paragraph and may cite a few sources. So the contest is no longer for position. It is for being included at all.
Google began showing AI Overviews in the United States in May 2024. OpenAI launched ChatGPT search at the end of October 2024. Both pull from web sources, and both change often, which is why anecdotes disagree. Treat claims warily.
Clicks fall when answers sit above the list. Search Console shows it. What is hard is knowing why a given brand is named, and nobody outside the vendors can see the weighting.
Measurement is the second problem, and it is harder. Search Console reports classic impressions and clicks. It does not tell you how often a chatbot named your brand. Third-party trackers sample prompts, so treat their numbers as rough. This is slow work. Make decisions on steady patterns, not on one screenshot.
Signals that plausibly matter
Plausible is the key word. These are reasonable hypotheses, not findings.
Being mentioned elsewhere
Models learn from text. A brand that appears in many independent pages is easier to name. Links are one visible form of that.
Clear, quotable pages
Short direct answers near the top, tables, definitions. Easy to lift, easy to cite.
First-hand detail
Specifics that generic text lacks: numbers, steps, screenshots of your own work.
Crawlable and fast
If bots cannot fetch the text, nobody can quote it.

Testing AI visibility without fooling yourself
Start with a prompt list, not a screenshot. Pick ten questions a buyer in your niche would type into a chatbot. Keep that list fixed for three months. Changing the questions every week makes any trend invisible. Hold the line on that rule.
Run each prompt in a clean session and record the answer. Note which brands appear and the order they appear in, since a reader often meets the first name first. Write the date beside every row.
Record the answer, not just a yes or no
Whether your brand appears is only half of the picture. Note how it is described, whether a link is offered and which competitor sits next to it. A mention framed as a warning tells you something a bare yes never will.
Keep the raw answer text in a file. A later review is easier when you can read what the model actually said, rather than your summary of it.
Expect noise from one-off results
Answers change between runs, between accounts and between days. One screenshot proves almost nothing. Run the set twice a month and look for patterns that hold across several runs.
If your brand appears in one answer out of ten, that is a fact worth tracking. Treat it as a baseline, not a victory.
Can a tracker do this for you?
Some third-party trackers sample prompts on your behalf. They save time, but they sample on their own schedule and with their own prompts. Check what they test before you trust the chart, and compare it with your own list once in a while.
Where a link API fits
It cannot tell you what a chatbot says. That limit is real. It can tell you who links to you and to the brands that do get named. Pull refdomains for both brands and look for linkers you lack.
Try intersect on two or three brands that appear in answers for your topic. Domains that link to all of them are the places the whole niche is known from. That is a practical outreach list. It holds whichever way AI search turns out.
Run the pull monthly and keep the files, because the point is trend, not a single reading. See the teardown steps for the exact calls.
A reasonable reading is that linking domains are an input you can see and improve. That is the honest part. Whether a model weighs them is not something anyone outside the labs can confirm. Do the work that holds under either answer: be useful, be cited, be easy to crawl. For wider context on cost and tooling, read what these tools cost.
What to do this quarter if you want more mentions
The honest plan has three parts, and none of them is a secret trick. Each part also helps classic rankings, which is useful when nobody knows how the next answer engine will behave.
Think of this as a small set of habits. You can keep them up whether or not you ever see your brand in a chatbot list, and they do not depend on any single vendor.
Make pages easy to quote
Put a direct answer near the top of the page, in one or two sentences. Follow it with a table or a short list that a reader can copy. Specific numbers help too, provided you can stand behind them and date them.
Write headings as questions or concrete claims. A heading such as what a link API returns is easier to lift than a clever title.
Earn mentions where your niche already talks
Mentions come from pages other people already read. Look for the sites that rivals are cited on, using the intersect call from the competitor teardown. Those are the places where your brand is likely to be named, or should be.
Offer something a writer can cite: a method, a small dataset, a worked calculation. Do not pitch a link for its own sake, because a placement nobody reads does not change how a model describes you.
Is this GEO or SEO? Usually both
The labels are a distraction for a small team. Generative engine optimisation and classic SEO share much of the same work: clear pages, real expertise and pages that load and can be crawled. Measuring ai search visibility is hard, but the content work helps either way.
If you can only do one thing this month, choose the page that answers a real question. Then read why rankings do not move to see what else could be holding it back.
Dates for context
AI Overviews in the US
Google began showing generated summaries above results.
ChatGPT search
OpenAI added web search with cited sources.
AI Mode
Google introduced a conversational search mode in the US.

Reading a quarter of AI results without overreacting
Three months of results give you something a single week cannot. Patience pays here. Patterns start to show, and so do the false alarms. Read the file as a sequence, not as a set of wins and losses.
Before you change anything on the site, decide what would count as a real shift. Write the threshold down first, so a bad afternoon does not set the plan.
Separate a new name from a lost one
When a competitor appears in answers and you do not, check whether the question changed before you check the brand. Many apparent losses come from a different prompt mix rather than from a different model.
A new name that shows up once is noise. A name that holds its place across three runs is a signal worth a teardown.
Watch the link side as a separate line
Keep the link counts in their own column, next to the mention counts. The link data moves on the monthly refresh, while your mention file may move every fortnight. Mixing them makes both harder to read.
If linking domains grow and mentions stay flat, the next step is outreach to pages that already discuss the topic. If mentions grow and links stay flat, look at what the new mentions cite.
When to stop and fix the page instead
If you cannot find your brand for any of the ten questions, the page itself may be the problem. A page that does not answer the question plainly gives a model little to quote, however many links it has.
Fix the page first, then wait a full cycle before judging the result. Changing the page and the prompts at the same time leaves you unable to tell which one mattered.
Do not compare across different models
Results from one assistant do not carry over to another. Each one draws on its own sources and updates on its own schedule. Keep a separate column per assistant, and avoid averaging them into one score that describes none of them well.
The same applies to regions and languages. A prompt run from one country can return a different set of names than the same prompt run elsewhere. Note the location with every row.
Questions we hear
Will brand mentions replace backlinks?
Unknown. A link is a mention that machines can count easily, so it is unlikely to vanish soon.
Is classic SEO dead?
No evidence of that. Most AI answers still draw on pages that earn visibility the usual way.
Can your API measure AI citations?
No. It covers link data only. Use it for the link side of the story, see the endpoint list.