Search stopped being a list of links and became an answer. That single change breaks the assumption underneath a decade of SEO tactics: that ranking produces a click. Today you can rank, be summarized, and never be visited — which means the strategy has to optimize for something other than position.
Two jobs that used to be one
Traditional SEO earned a position in a list. AI SEO — sometimes called answer engine or generative engine optimization — earns a citation inside a generated answer. They overlap heavily but they are no longer the same job, and the second one is where the growth is.
- ›Classic search — position determines clicks, and the click is the outcome you optimize.
- ›AI answers — being cited determines whether you exist at all, and the citation itself carries brand value even without a click.
- ›The overlap — crawlable, well-structured, factually specific pages still feed both. Nothing here asks you to abandon fundamentals.
The uncomfortable version of this shows up in your own analytics before anyone explains it to you. A page holds its position. Impressions hold steady or climb. Clicks flatten, then slide. Nothing is broken and no penalty landed — the question simply got answered before the reader reached the choice of where to click. Informational queries lose click-through first because they are the easiest to summarize. Transactional and local queries hold longer, because booking someone still requires leaving the answer.
There is a compensating gain, and it is easy to miss if you only watch session counts. The visitor who arrives after reading an AI answer has already been pre-briefed on what you do. Fewer sessions, warmer sessions. So watch conversion rate per session and branded search volume, not raw traffic. If sessions slide and booked calls hold, nothing is wrong. If both slide, something is.
What actually earns a citation
Language models assembling an answer pull toward sources that are unambiguous, specific, and easy to attribute. That has practical consequences for how a page is written, and it rewards a different style than link-chasing content did.
- 01Answer the question in the first two sentences. Burying the answer under an introduction hands the citation to whoever did not bury it.
- 02Be specific enough to be quotable — concrete numbers, named steps, real constraints. Hedged generic prose gets averaged away against every other hedged generic page.
- 03Structure with honest headings that match how people actually phrase the question, not how your industry phrases it internally.
- 04State your scope plainly: who this applies to, where, and when. Models favor sources that qualify themselves, because a qualified claim is safer to repeat.
- 05Keep facts current and dated. Stale specifics are worse than no specifics — a wrong number attached to your name is a liability, not a citation.
Write the answer, then the article
Take a real question: how much does an AI receptionist cost? The version that never gets cited opens with three paragraphs about how the industry is changing. The version that gets cited opens by naming the three billing models the category actually uses — per seat, per minute of handled call time, or a flat platform fee with a usage meter — says which one suits a small crew, and then explains the reasoning underneath. Same information. Different order. The order is the whole difference, because a model extracting an answer reads from the top and stops when it has one.
The unit of extraction is the passage, not the page. Write paragraphs that survive being lifted out alone. That means no "as mentioned above," no pronouns pointing three paragraphs back, and a named subject in the first clause. It reads slightly more repetitive to a human skimming top to bottom. It reads correctly to everything else, including the reader who lands mid-page from a search result.
Five tactics that still carry weight
Most of the AI SEO advice in circulation is fundamentals restated with new vocabulary. These five genuinely changed in emphasis rather than just in name.
- 01Topical depth over keyword breadth — a cluster that covers a subject completely gets cited more than scattered pages chasing separate terms, because the model is looking for one source that resolves the whole question.
- 02Entity clarity — make it unambiguous what your business is, where it operates, and what it does, in structured data and in plain prose.
- 03First-hand information — original numbers, real processes, and things only you could know are the hardest inputs for a model to synthesize from elsewhere, which is exactly why they get attributed.
- 04Machine-readable context — schema markup, clean HTML, and a current llms.txt so crawlers get facts rather than inference.
- 05Reputation surface — reviews, profiles, and third-party mentions, since answers about a business assemble from more than that business's own site.
Entity clarity, concretely
Entity clarity is not a philosophy, it is a chore. Write one canonical description of the business and use the identical wording on the site, the Google Business Profile, the directory listings, and the social bios. Match the business name, address, and phone number character for character across all of them — "Ste 4" in one place and "Suite 4" in another is a small inconsistency that costs you a confident match. Add LocalBusiness schema with a sameAs array pointing at every profile you control. Name your service area in prose, not only in markup, because prose is what gets read back.
First-hand information you already have
Service businesses usually think they have nothing original to publish, then describe a dispatch process no competitor has written down. You have call volume by hour, the share of jobs that come from a repeat customer, the seasonal shape of your demand, the four questions every caller asks before booking, and the real reason estimates go cold. That is first-hand data. Publish it as a range, label it as your own operating data rather than an industry benchmark, and date it. Honest and specific beats impressive and unsourced — an invented statistic is the fastest way to get repeated wrongly.
The tools, and what each one is actually for
Tools are instruments, not strategy. It still helps to know which instrument answers which question, because most of the confusion in this category comes from buying one that measures something you were not asking about.
- ›Ahrefs and Semrush — demand discovery and link context. Use them to find what people ask and who else answers it, not to chase a difficulty score.
- ›Google Search Console — the only ground truth you own for classic search. Compare the impression line against the click line per query group; that gap is the story.
- ›Clearscope, Surfer, and MarketMuse — coverage graders. They tell you which subtopics the existing top results cover and you do not.
- ›Screaming Frog — crawl and structure. It finds the broken headings, orphan pages, and missing markup that quietly keep a page out of consideration.
One honest limit on the coverage graders: they optimize toward the average of whatever currently ranks. That average is exactly the smoothed, hedged prose an answer engine skips over, because it adds nothing a model could not assemble from the other nine results. Use a coverage grade as a checklist for what you forgot, then deliberately write the part the tool cannot score — the specific number, the named constraint, the thing you learned by doing the work.
A newer category of tool tracks AI answers rather than rankings, and the billing unit tells you what it actually measures. These platforms meter on tracked prompts, projects, and audit volume rather than on words written — Writesonic, for example, now sells itself as an SEO and generative-engine platform priced this way. That unit is the product description: you are paying to watch a fixed set of questions and see whether your brand shows up in the answers. Buy the unit that matches your question, and read the tracked-prompt allowance on the tier you would actually purchase rather than the headline number, which usually applies only at the top tier. No vendor sees every answer, because answers vary by user, phrasing, and day. Treat any of these dashboards as a sample, not a census.
You can also do this for free. A spreadsheet with ten questions, three assistants, and a monthly date column is a legitimate measurement system, and it beats an expensive dashboard nobody opens.
Related: what llms.txt does and how to publish one
A sequence that fits a small team
AI SEO does not require a bigger content operation. It rewards a more concentrated one — depth on a narrow subject beats breadth across many, and that favors small teams that pick a lane.
- 01Pick one subject you can credibly own, narrower than feels comfortable. "Emergency HVAC repair in one metro" is a lane. "Home services" is not.
- 02Publish the definitive page on it, then the supporting pages that answer every adjacent question — pricing, timelines, what goes wrong, what it costs to wait.
- 03Fix the machine-readable layer once — schema, llms.txt, clean markup, consistent business details everywhere you are listed.
- 04Query the major AI assistants monthly with your customers' real questions and log whether you are cited, and whether what is said about you is correct.
- 05Expand only after you are consistently the cited source on the first subject. Two owned subjects beat nine contested ones.
The monthly thirty-minute check
Write down the ten questions a customer asks before they book — the real phrasing, including the awkward ones about price and timing. Ask each of the major assistants those questions on the same day each month, in a fresh session so prior conversation does not skew the answer. Record four things: were you named, was a competitor named, was the fact about you correct, and what source did the answer lean on. The fourth column is the useful one. It shows you where the answer is being assembled from, which is often a directory or a review profile rather than your own site — and that tells you where the next hour of work goes.
Two limits worth saying out loud. This is slow — expect a couple of months before a new page shows up in answers at all, because the models have to encounter it, and re-encounter it. And none of it is controllable the way a rank was: you can make yourself the most citable source on a subject and still get skipped on any given day. Aim for the trend line, not the individual answer.