AI search optimization services are meant to make your business appear when someone asks an AI assistant a question in your category. The honest ones do three things: measure whether you currently get named, fix the ordinary search fundamentals underneath, and build pages that answer real questions well. Most of what else gets sold in this category is either unnecessary or unproven, and Google has said so in writing. What is really working is quieter than the pitch, and we will get to the specific reason by the end.
One note on sourcing. Everything below is tied to either platform documentation from the companies that run these systems, or peer-reviewed and preprint research. We deliberately cite no marketing blogs here, because almost every published guide on this topic is written by a company selling the conclusion it reaches. Search volumes and difficulty scores come from our own keyword research, a pull of 1,341 healthcare keywords, July 2026.
Why this matters right now
The demand is real and it is new. In our own keyword data, “ai search optimization services” draws 480 searches a month at a difficulty of 8 out of 100, with a cost per click of $25.59. The part that matters more is the trend line. Indexed across 12 months, it starts at 0.05, climbs through the middle of the year, peaks at 1.00, and sits at 0.76 in the most recent month. A year ago almost nobody searched this. Now they do, and the people answering them are mostly vendors.
That combination, high commercial value and near-zero independent information, is how a category fills up with expensive nonsense. It is also why the buying decision is so hard right now. You are being asked to pay for something new, in a field where the measurement is immature, by people who benefit from you not asking too many questions.
So this article is written the way we would want it written for us. Here is what the platforms say, here is what the research shows, and here is what remains honestly unknown.
What gets sold that does not work
None of this is fraud exactly. Most of it is a plausible idea that got productized before anyone checked whether it worked, and then kept selling after the platforms said it did not.
The llms.txt file. This is the flagship deliverable across much of the category, a text file placed on your site to tell AI models how to read it. Google’s position is unusually direct: you do not need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search, and creating them will neither harm nor help your site’s visibility or rankings. Not “we prefer other signals.” Neither help nor harm.
Special AI schema. The same guidance says structured data is not required for generative AI search and there is no special schema.org markup you need to add. Schema remains worth having for other reasons. It is not the entry ticket it gets sold as.
Mass mentions and placements. The pitch is that AI engines trust third-party sources, so a service will place you across a set of sites. Google addresses this one too, noting that seeking inauthentic mentions across the web is not as helpful as it might seem. There is a real insight buried in the pitch, which is that independent coverage does appear to matter. The version sold as a package, identical placements bought in bulk, is not that.
Keyword stuffing, reborn. Some services still write pages dense with variations on the theory that models match on terms. The research is unkind here. Keyword stuffing reduces the position-adjusted visibility metric, and multiple benchmarks confirm null or negative effects. It is worse than doing nothing.
Guaranteed AI rankings. This is the one to walk away from. We will come back to why it cannot be promised.
The common thread is that all five sell certainty about a system nobody outside these companies can see inside. Google says as much itself, in a caution box in its own guidance: be wary of third-party tools that promise ranking success or claim to use “internal” Google metrics, because no third-party tool has access to its internal ranking or AI systems.
What the evidence supports
Now the useful half. A 2026 academic survey reviewing three years of generative engine optimization research, published on arXiv, sorted the proposed techniques by how well they hold up under controlled testing. Four things survived.
| What | How strong | What it means for you |
|---|---|---|
| Answering the actual question | Strongest and most reproducible | Models favor alignment with the question over credibility signals |
| Position in retrieval | Beats most rewrites | Being findable in the first place outranks wordsmithing |
| Extractable evidence | Consistent gains | Statistics, definitions, comparisons, dates, references |
| Recent dates | Moderate | Mostly on time-sensitive questions |
Answer the question that was asked. The survey found query relevance the most reproducible lever of all, and noted something worth sitting with: models favor alignment with the question over credibility signals. A page that plainly answers the thing someone asked beats a more authoritative page that circles it. Most business websites are written to impress rather than to answer, which is why so many of them never get quoted.
Get retrieved before you get clever. Position in the retrieval context turned out to matter more than most text rewrites. That sounds abstract until you notice what it implies: ordinary search visibility is upstream of everything. Google says this outright: SEO best practices stay relevant because its generative AI features are rooted in its core Search ranking and quality systems. And to be shown as a supporting link in AI Overviews or AI Mode, a page must be indexed and eligible to be shown in Google Search with a snippet. There are no additional technical requirements.
Put quotable facts on the page. Statistics, definitions, comparisons, prices, dates, references. The survey reports quotation addition producing around a 41% relative gain in the foundational study. Engines quote what is quotable. Vague pages give them nothing to lift.
Then the counterintuitive part. In end-to-end testing, optimizing page bodies alone reduced average top-20 presence by about 9%, top-10 presence after reranking by 16%, and final citation by 6%. Read that again, because it is the single most useful finding in the literature. Pages tuned for AI citation in isolation did worse overall, because the tuning damaged the retrieval that would have put them in front of the model to begin with.
That is the mechanism behind every “we optimized for AI and traffic dropped” story we have heard.
The one thing that is checkable
Most of this field resists verification. One piece does not.
ChatGPT’s appearance in search is governed by a crawler called OAI-SearchBot, and it is separate from GPTBot, which governs training data. You can allow one and block the other. If your robots.txt blocks OAI-SearchBot, or your host or CDN blocks the published crawler traffic, you will not be shown in ChatGPT’s search answers, though you can still turn up as a navigational link. No amount of content work changes that.
We check this first in every visibility audit we run, and we find it broken more often than you would expect. Usually nobody chose it. A security plugin tightened the rules, or someone blocked a list of bots years ago and the list grew. It is free to fix and it takes minutes.
What to do in the next two weeks
- Check your crawler access. Read your own robots.txt, confirm OAI-SearchBot is not blocked, and confirm your CDN is not blocking it either. About 30 minutes.
- Open the free report. Google Search Console includes a generative AI performance report covering Google’s AI surfaces. You already pay nothing for it. About 20 minutes.
- Ask the engines about your own market. Type the questions your customers ask into three or four assistants, plainly, and write down who gets named. This is the baseline every paid service should be giving you, and you can do a rough version yourself in an afternoon. About 2 hours.
- Fix the answers on your own pages. For your top five questions, put the answer in the first two sentences of the page. Not the setup, the answer. About 1 hour per page.
- Add the quotable specifics. Prices, timeframes, comparisons, dates. Anything a model could lift as a sentence and attribute to you. About 2 hours.
Total is roughly a day and a half of real work, spread across two weeks, and it costs nothing. Steps 1 and 2 are the ones nobody does, and they are the ones with the clearest evidence behind them.
So what’s really working?
Here is the part the category will not tell you.
The 2026 survey’s overall conclusion is that the evidence is strong for a causal effect only once your content has already been retrieved, moderate for certain content properties, and weak for whether any of it flows through to traffic. Its central finding is one sentence: no reviewed technique shows a stable, longitudinal, cross-platform causal effect on organic discoverability.
Nobody has shown they can reliably get you discovered. Everything demonstrated so far operates on content the engine already found.
And the engines do not even agree with each other. Research comparing them found roughly a quarter domain overlap between two major surfaces, with similarity below 0.2 in some comparisons. There is no single AI search to be ranked on. There are several separate systems with separate sources.
So what is really working in 2026 is not a technique. It is a position. Be findable through ordinary search, because that is the gate everything else runs through. Answer questions directly, because relevance beats authority in these systems. Put quotable facts on the page, because engines quote what can be quoted. Then measure what the engines say about you, repeatedly, because measurement is the only part of this field that is real right now.
That last one is the whole game. In a category where the mechanism is uncertain, the vendor who shows you the raw answers is worth more than the vendor with the confident framework. Ask for the evidence. If they cannot show you what the engines said before they started, they cannot show you what changed.
The services are just how you get there, and quite a lot of them are selling you a file that does nothing.