Every enterprise SEO team we work with in September 2026 is running the same experiment. They are trying to figure out which of their content investments actually earn citations in ChatGPT, AI Mode, Perplexity and Copilot, and which just fill a calendar. Most of them are running the experiment in the dark, because the loop between publishing and being cited is longer than a Search Console impression and lower fidelity than a rank tracker. This playbook is the shape of the work that closes that loop.
It is not a checklist of tactics. There are already enough of those. This is five plays, each with the concrete moves that make it work and the criterion that says it is done. Every play maps to a specific failure mode we see in enterprise engagements: the pages that read well but do not answer the question, the entity graph that names your competitor when the engine paraphrases you, the comparison surface a first-party site refuses to build, the measurement stack that reports impressions instead of appearances, and the monthly review that never happens.
How to read this playbook
Each play is written to be run once by an owner, then folded into a monthly cadence by that same owner. A play is done when its acceptance criteria are met and evidence is filed, not when the ticket is closed. If you have shipped a play but cannot produce the evidence, the play is still open. That rule is the entire reason the playbook works.
The five plays are ordered on purpose. Plays 1 and 2 are foundational: the answerable page and the entity mirror. Plays 3 and 4 are competitive: the comparison surface and the instrumentation that lets you defend it. Play 5 is the cadence that keeps the first four from decaying, which is the play most teams skip and then wonder why the results reverted.
Play 1: Fix the answerable page before you fix anything else
An answerable page is one whose top viewport, in the raw HTML that a scraper sees, contains one clear answer to the question the URL implies. Nothing on the modern enterprise site is that shape by accident. Homepages open with a value proposition. Product pages open with a hero animation. Blog posts open with a preamble. All three are bad for citations, and none of them read as broken to a human.
The fix is not to rewrite everything. It is to identify the twenty to fifty pages that are the load-bearing surfaces for your category, and rewrite the top of each of them so a scraper that only sees the first two thousand tokens finishes with a specific answer, not a headline. This is dull, high-leverage work, and it is what we do first in every generative engine optimization engagement.
The reason the raw HTML matters is that most AI engines run their content fetches with lightweight fetchers, not full browsers, and the crawler does not sit through a hero animation. Even the engines that render JavaScript do not wait for it the way a human does. The page shape that wins citations is the page shape a text-mode fetcher can answer from. This is the same reason a lot of SPA-heavy sites appear invisible in AI search despite ranking fine in Google, and it is a variant of the failure mode we walked through in the Googlebot HTTP method firewall blocking teardown.
Two adjacent moves make this play stick. First, add a small block near the top of every answerable page (a callout, a stat band, whatever fits) that contains the two or three numbers a comparison prompt would pull. Engines quote numbers when numbers are present in a scannable shape. Second, put a real 'last updated' date next to that block. Engines and readers both weight recency more than teams believe, and a stale date is the cheapest way to lose a citation to a competitor whose page is functionally identical but three months newer.
Play 2: Own the entity in the third-party mirror
The engines paraphrase what the web says about you, not what you say about yourself. When ChatGPT describes your product in a category question, it is not summarizing your homepage. It is summarizing Wikidata, G2, Peer Insights, Reddit threads, and a handful of trade-press pieces. If those sources say the wrong thing, your homepage cannot fix it.
This play is unglamorous and one of the highest-leverage things you can do in a quarter. It is also the work that PR and SEO both assume the other owns and neither actually does, which is why the entity mirror at most enterprise brands is stale and wrong.
The reason this play matters more every quarter is that the engines are increasingly co-training on their own citation graph. A wrong description in a Wikidata property that got scraped in August 2026 is a wrong description that shows up in an answer generated in November, even if you fixed the Wikidata property in September, because the model has already seen the earlier version. Fixing this early is not optional. It is what compounds.
The failure mode to avoid here is treating the mirror like a brand exercise. Nobody at the engine cares that your marketing team calls the category 'unified endpoint observability' if the analyst pages, the Reddit threads, and the buyer prompts all call it 'endpoint monitoring'. Alignment beats aspiration. If the mirror uses the older category name, meet it there in your own copy first, then move both together.
Play 3: Publish the comparison layer the engines cannot
Comparison prompts are where AI answers most often override the first-party site. 'Best X for Y', 'X vs Z', 'alternatives to A' all get answered by a paragraph that aggregates across G2, Reddit, and a handful of comparison articles most of which are written by aggregator sites for lead-gen. Those aggregators are the citation of record for these queries almost by default, and it is because most enterprise vendors refuse to build the comparison surface on their own site.
The correct response is not to complain about it. It is to publish the comparison surface, with the aggregators' scoring rubric, but with data you can defend. The engines will cite the honest first-party comparison over the aggregator because it has more specific data. The reluctance to build one is a self-inflicted wound.
| PAGE TYPE | WHO IS WINNING TODAY | WHY THEY ARE WINNING | WHAT TO BUILD |
|---|---|---|---|
| 'X vs Y' pages | Aggregators (G2, Capterra, Sourceforge) | They have both products' feature matrices in one place, in machine-readable shape | A fair-frame comparison on your site, real feature checklists, price bands where legally allowed, a note where the competitor beats you |
| 'Best X for Y' pages | Listicles written by SEO agencies | They exist; nobody else built one | A category page that names three or four alternatives, honestly ranks them for specific use cases, and links to their sites |
| 'Alternatives to X' pages | Direct competitors bidding on your brand | The engines are pulling brand-plus-alternatives queries from competitors' pages that name you | Your own alternatives page for the products you actually replace, with concrete migration guidance |
| Category definition pages | Analyst firms (Gartner, Forrester) | They own the category vocabulary and the engines respect the source | A definition page that adopts the analyst vocabulary verbatim, cites them, and adds one unique frame nobody else has published |
The internal objection to this play is always the same: 'we do not want to promote competitors on our own site.' The counter is measurable. If the query is being answered by an engine anyway, and the engine is choosing between an aggregator's version and yours, the aggregator's version will always frame your competitor's strengths as a legitimate finding. Your version can frame them honestly and still name where you are the better choice. The one that never appears is the one that gets no share of that conversation. This is the same argument we made in the comparison decision table for link building.
Play 4: Instrument citations as a first-class channel
Most enterprise teams still measure AI presence with a rank-tracker analog: a dashboard that shows how often you 'appear' in some engine's citations, averaged across a fuzzy prompt list, updated on an unspecified cadence. That is the AI-native version of a vanity metric. It moves, but no decision comes out of it, and no owner knows why the number changed.
A real citation panel is small, controlled, and repeatable. Sixty prompts, four engines, one owner, one monthly cadence. That gives you 240 answer instances a month, which is enough signal to see real movement and short enough to run in half a day. Everything above that scale is nice-to-have. Everything below it is noise.
The three states of a citation panel prompt (illustrative distribution on a typical baseline)
The three-state split matters because the responses to each are different. A 'cited-competitor' result is a comparison-surface problem (Play 3) or an entity-mirror problem (Play 2). A 'cited-nobody' result is an answerable-page problem (Play 1) or a category-definition gap. A 'cited-you' result is the win state, and the next question is whether the URL cited is the right URL, not whether the citation exists at all. Half the enterprises we onboard have their homepage cited when their pricing page would have been the higher-leverage answer, and the fix is a page-level content decision, not a channel-level one.
The reason we insist on three months of consecutive runs is that the first month is the baseline, the second month is the noise floor, and the third month is the first time you can talk about a trend. Skipping a month resets the clock, and the noise from a two-month gap eats the signal. We wrote about the same drift problem in the error bar framework for AI search reporting and it applies here too: monthly is the minimum, not the aim.
The panel is also the input that lets you make budget decisions. When a stakeholder asks whether the GEO work is paying off, the honest answer is the panel movement across the three months, tagged by root cause, with a specific example of a prompt whose answer improved because of a specific page you shipped. That is not a dashboard vibe. It is a defensible number, and it is what the reporting and analytics work we run for clients revolves around.
Play 5: Run a monthly citation-loss review
The plays above are individually powerful and collectively pointless without a monthly review that closes the loop. This is the play most teams skip, because it is the play with no ship date. It is the review that decides what to ship next, and it is what turns four one-time projects into a system.
The review is a single meeting, thirty minutes, once a month, run on the day the citation panel is published. Three losses are picked, three root causes are named, and three fixes are scheduled. That is the whole meeting. No decks, no summary, no committee.
The reason six months is the acceptance criterion is that anything shorter is a project, and anything longer waits too long to prove the loop works. A team that runs the review six times in a row has built a system. A team that runs it twice and then cancels the third has produced a report. There is no in-between.
The failure mode inside the review is the same one in every quarterly business review at every enterprise we have worked with: the meeting becomes a status update instead of a decision meeting. Guard against it explicitly. The output of the review is a written list of three fixes, three owners, three ship dates. If the output is a set of slides, the review has failed and the next one should reset to the three-fix format. This is the same discipline that makes our audits generate change instead of paper.
How the five plays fit together
The plays are not independent tracks. Play 4 (the panel) is the instrument that says which pages need Play 1 work, which entity mirrors need Play 2 work, and which comparison surfaces need Play 3 work. Play 5 (the review) is the cadence that turns the panel results into shipped fixes. Without the panel, the other plays run blind. Without the review, the panel becomes a dashboard nobody reads.
The economic argument to make internally is that the plays compound with each other. Fixing a page (Play 1) does not just improve that page's citation odds, it improves the entity paraphrase (Play 2), because your own page is one of the sources the engines pull. A better comparison surface (Play 3) makes the citation panel (Play 4) resolve faster because there is a clear winning URL to be cited. A monthly review (Play 5) makes all of the above self-correcting. The plays run individually would still work. They run together at a materially better rate, and the difference shows up in the panel by month four or five.
One last thing about scope. This is a playbook for the answer surface, not the whole marketing stack. If your organic search program is broken, fix that first; AI citations correlate with the surfaces that already rank well, and no amount of playbook work compensates for a site that Google will not crawl or a content team that ships three posts a quarter. If your product is not actually differentiated in the ways your marketing claims, the fair-frame comparison in Play 3 will be honest about that, which is a product problem for a different team. The plays are what a strong practice does inside a strong foundation, not a substitute for either.
Run the five plays. Publish the panel every month. Ship the three fixes before the next review. Do that for six months and you will have moved the citation-competitor split by twenty to forty percentage points on a real prompt panel, which is the number that shows up in the enterprise B2B engagements we finish this way. Do it for a year and you will have built the citation practice that most competitors are still talking about instead of running.
See where you are cited today
A free snapshot audit of your rankings and AI citations before we ever talk.
Tyler leads work at the intersection of SEO and generative engines at Something Inc., helping B2B brands get ranked and cited across every major AI engine.