AI search keeps changing. Who is tracking it for you?
The engines change what they retrieve, how many sources they cite and how they attribute them, and they did it more than once this year. Nothing warns you. There is no alert, no error and usually no traffic drop, because a large share of the value was never traffic in the first place. So a setup built for how things worked last quarter keeps looking fine while it stops working, and the decline gets noticed months later and blamed on something else.
By Liron Segev. Last updated August 2026.
Four things that shift without telling you
What gets retrieved
Whether an engine reaches for live information at all, and how aggressively, varies by question type and shifts as the products get updated. A category resolved from memory one quarter can be resolved from live retrieval the next.
How many sources get cited
The difference between an engine citing two sources and ten is the difference between winner takes all and being one of several. Those counts are not fixed, and a change moves everyone's position at once.
How attribution is displayed
Whether a business is named in the text, linked as a reference, or both, decides whether a buyer ever registers you. The same underlying result can be worth a lot or nothing depending on how it is surfaced.
Which engines people actually use
Where the questions get asked moves faster than anything else here. Optimising for the assistant your buyers used last year is a very expensive way to be thorough.
The decay is silent, which is the whole problem
If your website went down you would know within an hour. If your ads stopped delivering you would see it the same day. Both of those failures announce themselves, which is why neither is particularly dangerous.
This one does not. Your pages still load. Your rankings may be unchanged. Your traffic looks normal, and would look normal even if you lost every mention you had, because being named in an answer often produces no click at all. Every dashboard you own can be green while the thing you actually bought quietly stops happening.
The failure shows up somewhere much less useful. Fewer inbound conversations. Callers who have not heard of you. A competitor being mentioned in a meeting. By then the cause is months old and it looks like a market problem rather than a fixable one.
That is the actual argument for measuring on a schedule. Not because measurement is valuable in itself, but because this is a category of failure with no natural alarm.
Three honest choices
Ignore it
Defensible for a while. Published answers keep working for months after anything changes, so nothing appears to break. The cost arrives as a slow decline nobody attributes to a cause, usually noticed when someone runs a check for an unrelated reason.
Fine short term, expensive quietly
Do it yourself
Genuinely possible. Run your category's questions through the engines monthly, keep notes, notice when the shape of the answers changes, and adjust. Two or three hours a month if you are disciplined about it, and more when something actually moves.
Works if you will really do it
Pay someone to watch
Only worth it if watching is attached to changing things. A report telling you what shifted, with nothing rebuilt in response, is a subscription to information you cannot act on.
Worth it when it includes the rebuild
The middle option is real and we would rather say so. Plenty of owners can run ten questions through four engines once a month and act on what they see. If that is you, do that. The reason most people end up not doing it is not difficulty, it is that it is the easiest thing on a list to skip in a busy month, and skipping it produces no immediate consequence at all.
How to tell watching from billing
If you are paying someone for ongoing work, the question that separates a real retainer from a subscription is simple: what got rebuilt last month, and why.
A report is not the deliverable. A score is not the deliverable. Being told your visibility score moved a few points in either direction is only useful if it is attached to a decision about what to change. Plenty of services in this category have quietly become monitoring products with content attached, and they are priced like the work.
Ask for the last three months of changes made in response to something observed. If the answer is a list of articles published on a schedule that would have been the same regardless of what the engines did, the watching is decorative.
This is a fair question to ask us, and it pairs with the other one worth asking anybody: what do you keep if you leave. Together they cover most of the ways this goes wrong.
Why this is the retainer and not an upsell
We rebuild against what the engines currently do rather than against what they did when a client signed. Questions get re-run across engines, new gaps get found, answers that have stopped performing get reworked, and new buyer questions get added as categories shift.
That is the entire justification for an ongoing fee, and it is why the build and the retainer are priced separately. The build is a one-time job. Staying present is not, and pretending otherwise would mean selling a finished product that quietly expires.
The client's system stays in their own environment throughout, so the work accumulates somewhere they keep. If they stop, they stop with a current system rather than a stale one.
Related questions
How often does AI search change?
Continuously, with several noticeable shifts a year. The engines update what they retrieve, how many sources they cite and how they present attribution, and none of it is announced in a way a business owner would ever see. The practical effect is that an approach built around how one engine behaved last quarter degrades without any visible event.
Will I know when my AI visibility drops?
Not unless you are measuring on a schedule. There is no alert, no error and no traffic cliff, because a large share of the value was never traffic in the first place. A business can lose most of its mentions and see nothing at all in its analytics, which is why the decline is usually discovered months late and attributed to the wrong cause.
Do I need to keep paying after the content is published?
You need someone to keep watching and adjusting, which may or may not be a vendor. Published answers do not stop existing, but the conditions they were built for move, competitors publish answers to the same questions, and new questions appear as your category changes. Content is a one-time cost; staying present is not.
Can I just check this myself?
Yes, and for many businesses that is the right answer. Run your ten most important buyer questions through ChatGPT, Claude, Gemini and Perplexity once a month, record who gets named, and watch for the pattern changing rather than for a single bad result. The work is not difficult. It is just easy to stop doing, and stopping is the failure mode.
What actually decays if nobody watches?
Position on questions you already own, as competitors publish their own answers. Coverage of new questions that did not exist when your content was written. And the structural fit between how your content is built and what the engines currently reward. The third is the slowest and the most expensive to fix late, because it usually means reworking rather than adding.
Is this just a reason to charge a retainer?
It is the honest reason for one, and it is worth testing rather than accepting. Ask what specifically gets done each month and what changes as a result. If the answer is a report, you are paying for information. If the answer is that questions get re-run, gaps get found and content gets rebuilt against what the engines currently do, you are paying for work.
Other questions buyers ask
Find out where you actually show up
The free scan runs your business against ChatGPT, Claude, Gemini and Perplexity and shows you which buyer questions you appear in and which ones a competitor owns.