Start with the uncomfortable study, because it saves you two weeks. C-SEO Bench, an academic benchmark for conversational SEO tricks, tested the popular rewrite hacks across tasks, domains, and adoption rates. Most methods were largely ineffective and some actively hurt document ranking, while the boring basics held up: factual pages, real sources, and earned authority. The paper is arXiv:2506.11097, and its multi-actor experiment matters most: even methods that worked for one site stopped working once competitors adopted them too.

That finding sets up this plan. You cannot shortcut your way into AI answers, but you can measure where you stand, remove technical blockers, study which sources engines already quote, publish pages that state facts with evidence, and retest with the same questions. The field evidence points the same way: a 260-million-prompt study by Ahrefs found branded web mentions correlate with AI visibility more strongly than any other factor it measured, and the team behind the GEO paper (arXiv:2311.09735) measured up to 40% higher visibility in controlled runs when pages added quotations, statistics, and cited sources.

This is a 30 to 60 minutes a day plan. Fourteen days, six phases, one fixed question set that never changes. Pair it with the checklist page, which lists every step with what to do, how to verify it, and which tool to open. Nothing here promises citations, traffic, or leads. It produces dated measurements you can act on.

Measure Foundation Sources Content Distribute Retest Days 1-2 Days 3-4 Days 5-7 Days 8-11 Days 12-13 Day 14 Same question set at both ends. Dated tables only.
Fig. 1. The 14-day loop: six phases, one fixed question set. Days 1 to 2 measure, days 3 to 4 fix retrieval, days 5 to 7 map sources, days 8 to 11 publish facts, days 12 to 13 distribute, day 14 retests with the same questions.

Three ground rules

The plan only means something if you follow these.

  • Measure before you change. Day 2 produces your baseline. Every later number compares against it.
  • Change one thing at a time. If you rewrite three pages and edit your robots.txt in the same week, a move in the numbers tells you nothing.
  • Keep the number even when it is zero. A baseline of 0 out of 12 mentions is data. Engine model updates will move results you do not control, which is exactly why every record carries a date.

Days 1 to 2: write a fixed question set and take a baseline

The question set is the measurement instrument. Write 8 to 12 questions a real buyer would ask when they hit the problem you solve. Three principles, no fixed template:

  • Ask it the way a customer would. Plain words, the vocabulary from your support inbox and sales calls, not your product marketing.
  • Mix the three shapes buyers use: recommendation ("which tool should I use for X"), comparison ("A versus B for X"), and problem phrasing ("how do I fix X").
  • Split branded and unbranded. Half the questions name your product or company, half do not. The unbranded half tests whether engines name you at all; the branded half tests what they say about you.

On day 2, run every question through ChatGPT, Perplexity, and Gemini, one fresh chat per question, and fill the same three columns for each answer: named, yes or no; described accurately, wrongly, or not at all; which sources the answer cited. Screenshot the answers and save the table with the date.

For scale: the search surface behind this keeps shrinking. SparkToro measured 68.01% of US Google searches ending without a click in early 2026, and Ahrefs measured position-one click-through rate on keywords that trigger AI Overviews falling from 7.3% to 1.6% between December 2023 and December 2025. Customers who stopped clicking are asking somewhere else.

Days 3 to 4: run a free technical scan and fix the foundation

Answer engines retrieve pages before they cite them. If your site is unreachable, blocked, or a duplicate of itself, no content work lands. Start with the free scan: seven technical checks on your homepage covering reachability, meta basics, and crawler access. It runs in the browser and needs no account.

Fix what it flags, in this order: pages return a working response and load reasonably; site:yourdomain.com actually shows your pages in Google; one canonical address per page; robots.txt does not accidentally block crawlers you want. Blocking AI bots can be a deliberate business decision, and if you make it, make it knowingly (see the numbers in our breakdown of robots.txt AI-training tokens).

One deliberate non-task: llms.txt. Ahrefs' log study across 137,210 domains found 97% of published llms.txt files were never fetched even once in May 2026, and Google states Search ignores the file. Write one if you enjoy it; do not spend a day on it. The full evidence sits in our llms.txt adoption analysis.

Days 5 to 7: study the sources answers actually quote

Rerun your question set and, this time, collect the sources: every domain the answers cite, across all questions and all three engines. Sort them into buckets, for example industry publications, review platforms, business directories, community threads, competitor comparison pages, and your own site.

Two measured facts make this exercise worth three days. First, the citation pool is wider than Google's first page: across ChatGPT, Perplexity, and AI Overviews, only about 12% of cited URLs rank in Google's top 10 for the original prompt. Second, overlap is loosening: for Google AI Overviews, 38% of cited pages also rank in the top 10, down from 76% a year earlier, because Google now pulls from more refined queries (Ahrefs, 4 million AI Overview URLs).

Your output on day 7: two or three source types that recur across your answers and are realistic for you to appear in. That list drives days 8 to 13.

Days 8 to 11: publish facts-first content engines can quote

Write pages that answer one question each, state specifics, and cite their sources. The controlled evidence says specifics travel: the GEO study measured up to 40% higher visibility in generative engine responses when pages added quotations, statistics, and citations to their claims. That was a controlled academic run, not a market guarantee, but the direction matches what answers actually cite.

The four page shapes that earn their keep:

  • A comparison page naming real alternatives, with your specific differences stated as facts.
  • An FAQ page answering the exact questions from your day 1 set, in the buyers' words.
  • A how-to page with steps someone can test without contacting you.
  • A numbers page covering pricing, benchmarks, or methodology, because answers need concrete figures to quote.

While publishing, sweep your entity facts: identical name, one-line description, services, and location across your site, profiles, and listings. Engines reconcile entities across sources; contradictions make them hedge or pick the competitor's version.

Days 12 to 13: show up in two or three places your buyers already read

The 260-million-prompt correlation study puts branded web mentions at the top of the factors it measured. Correlation, not proven causation, but the action it suggests is not spam: be present, with substance, where your buyers and your industry already are.

Pick two or three channels from your days 5 to 7 source buckets, not ten. Contribute answers that would still be worth reading if AI never quoted them. Get listed or reviewed on the platforms that recurred in your answers. Keep every fact identical everywhere. Then stop. Distribution beyond your real channels is where the tricks from the C-SEO Bench study creep back in.

Day 14: retest with the same questions and compare

Same question set, same three engines, fresh chats, same three columns. Fill table two and put it next to table one: mention rate per engine, accuracy of the descriptions, and which sources appeared.

Three outcomes, all useful. Mentions moved: keep the content phase running and retest in another two weeks. Accuracy moved without mentions: fix entity facts and profiles first. Nothing moved: that is a result too. Re-run the technical scan, tighten one page at a time, and keep the dates. Engine-side model updates will move numbers you did not touch, which is why the dated log matters more than any single reading.

What this plan will not do

It will not promise you citations, rankings, traffic, or leads, and anyone who promises that is selling past the evidence. The honest state of measurement today is three-legged: correlation studies (Ahrefs), controlled academic runs (the GEO paper), and a benchmark showing popular tricks fail at scale (C-SEO Bench). Displacement is real too: C-SEO Bench's multi-actor experiment found that when every competitor adopts the same tactic, the gains cancel. What the plan produces is a dated, honest baseline and a repeatable loop. If you want the deeper version, CiteGap runs 30 buyer-style questions across three engines; the written diagnosis reports what engines say about you and which sources they lean on, with no promises attached.

Run the free scan today, keep the checklist open beside you for the next two weeks, and judge the plan the way it asks you to judge everything else: same questions, dated numbers, one change at a time.

Free scan

Frequently asked questions

How much time does each day take?

Plan for 30 to 60 minutes. The day 2 baseline runs are the longest block, roughly an hour for 8 to 12 questions across three engines. Most other days are one scan, one page, or one contribution.

Do I need paid tools for this?

No. The engines have free tiers, the technical scan is free, and Search Console is free. Paid diagnosis buys depth and a second pair of eyes, not access to a different mechanism.

Which engines should I test?

ChatGPT, Perplexity, and Gemini are the default three because they answer with citations and can reach the live web. Add whatever your regional buyers actually use; the principle is the same fixed set every time.

Is llms.txt part of the plan?

No. Server logs across 137,210 domains showed 97% of published llms.txt files were never fetched in May 2026, and Google Search ignores the file. It is optional and costs ten minutes, so the plan spends that time on retrieval and content instead.

What if my brand is never mentioned at all?

Then your baseline is 0 out of 12, which is still a measurement and more than most businesses have. Follow the order: retrieval first (days 3 to 4), sources second (days 5 to 7), facts-first pages third (days 8 to 11). Mentions in unbranded questions are the slowest thing to move; judge on dates, not days.

Sources