AI search visibility engine (Airefs)

Wire the layers in reverse order of accuracy, the lead question first, because it is the only one that is actually true.

1. Lead pipeline (most accurate). Get a "How did you hear about us?" question into the signup or onboarding flow, and onto the discovery-call script. Name ChatGPT, Perplexity, Google, Reddit and word of mouth as options, plus a free-text box. Write the answer to a lead-source property in HubSpot so it reports next to every other channel.

<!-- Last-click badly undercounts this channel: a buyer discovers the brand inside an AI answer, then types the brand name into Google. It lands as branded or direct search and the AI never gets the credit. The self-reported answer is the only clean read, users who found you via an assistant know it and will say so. -->

1. Clicks (directional). Add the Airefs client-side script, a GA-style snippet, to catch human visits arriving from an AI answer. Read it as a trend only, never as a total.
2. Impressions (AI crawler activity). Needs server-side logs, because crawlers rarely execute JavaScript. Route by stack:

- WordPress, install the Airefs plugin. No code.

- Cloudflare, a lightweight Worker on any plan, or Logpush on Enterprise.

- Vercel / Netlify, a zero-code Log Drain on the paid tiers, or a small middleware / Edge Function on any plan.

- CloudFront, Fastly, Akamai, stream access or edge logs to the ingestion endpoint.

- Anything else, the server-side API call, or batch-forward via the HTTP Log API.

- Webflow or Framer, these do not expose server logs, so impressions cannot be tracked directly. The only route is a CDN in front: if the domain is proxied through Cloudflare, crawlers hit Cloudflare before the host, so integrate there. If it is not, escalate, say plainly that impressions are unavailable on this stack and run the system on prompts plus leads.

1. In the Impressions filter, switch on "User searches only" so the view isolates crawls answering live queries from background training crawls. Only the live ones are a visibility signal.
2. Human approval, confirm who owns the site changes before anything touches the client's stack.

Done when:

- the how-did-you-hear question is live and writing to a CRM property, and

- the clicks script is firing, and

- impressions are arriving, or it is documented why they cannot be.

1. In Airefs, pull the week's run: mention rate, rank when mentioned (first named is a different outcome from fifth), citations, and the competitor comparison.
2. Read every metric by tag, intent, segment, country. A blended number hides the only thing worth knowing.
3. Work the smart filters in this order, because that is their order of actionability:

- Recently lost, we were named, now we are not. Usually a competitor published, or a cited source changed. Act here first, before they cement it.

- Recently gained, proof something we did landed. Note what preceded it.

- Brand absent, the opportunity list: prompts competitors win where we are invisible.

- Brand present, confirm and hold.

1. Open Sources and read the URLs driving answers across the category. Two things matter: which format keeps getting cited (listicle, comparison, review roundup), and which top-cited pages look weak, stale, thin, years old, and are therefore displaceable.
2. Record mention rate, citation rate and share of voice against this system's metrics and note the week-over-week deltas.
3. Split the week's gaps three ways and hand off: content gaps to the article step, commentable threads to the discussion step, cited domains to the placement step.

Done when:

- the week's numbers are recorded, and

- the gaps are split across the three levers.

<!-- AI answers are probabilistic, trust the weekly aggregate, not day-to-day swings. Recency is a real lever: on recommendation prompts the model almost always runs a live search, so a genuinely new source can be picked up within days of being indexed. -->

1. Open the Articles backlog in Airefs. It pools gaps from five signals, prompt citations, SEO keywords, best practices, competitor content, and manual adds, so filter by tab rather than reading it as one pile.
2. Choose against the format evidence from the Sources read, not taste. If comparisons and listicles are what get cited in this category, that is what to write.
3. Queue up to five in Scheduled articles. Reorder, swap, or edit a slot's title, content type and extra instructions before it is written. Plan allotment: Lite 1/month, Pro 3, Expert 6.
4. The AEO agent drafts each one from the account Knowledge base, brand, competitive positioning, audience, writing style. Fill that in properly first: a thin knowledge base is the main reason a draft comes back wrong.
5. Review every draft against the citation bar before it goes anywhere:

- question-shaped headings that mirror real prompts, with the answer in the opening lines of each section;

- structured comparisons, tables, numbered lists, clear trade-offs;

- a neutral reference tone, not a sales page;

- dense and specific, named numbers, dates, sources;

- the brand positioned where it genuinely fits, not just a list of competitors.

1. Publish. Output is markdown, download or copy-paste. There is no CMS integration yet, so it lands in the CMS by hand.
2. It does not have to live on the client's domain to become a source. LinkedIn articles, Substack, Medium and partner sites all get cited.
3. Human approval, editorial sign-off before publish. Nothing ships as raw AI output.

Done when:

- the piece is live and crawlable, and

- it is logged against the gap it was written for.

1. Open Discussions in Airefs, commentable URLs the models are already citing for our prompts (Reddit threads, LinkedIn posts, Medium articles), plus new threads from Reddit Alerts.
2. Filter by Competitive gap first: threads where a competitor is named and we are not. Highest leverage in the whole system, no content production, and movement can show in prompt data within days.
3. Set Reddit Alerts on category terms, the brand name and top competitors. Timing is most of the value: an early comment on a forming thread shapes it, while a thread with hundreds of comments and a settled dynamic will not move.
4. Read the whole thread, then use Draft reply with AI to generate a comment from the thread plus brand context. Edit it until it sounds like a person, then post manually, Airefs does not post on your behalf.
5. The comment has to earn its place: answer the question, correct a misconception, or share a specific experience. Name the brand only where it genuinely fits. Promotional comments get downvoted or removed and the effort is wasted.
6. Which account posts is the client's decision, personal, team or brand. Agree it up front. Read the Reddit marketing guide on karma, content quality score, flagging and moderation before the first comment: a ban costs far more than the thread was worth.
7. Mark done or dismiss to keep the queue clean, and note which threads were seeded so the report step can check whether they surface in answers.
8. escalate if an account is flagged or shadowbanned, or a moderator removes a comment, stop posting in that subreddit and raise it rather than trying again.

Done when:

- the week's competitive-gap threads are engaged or dismissed, and

- seeded threads are noted for the report step.

1. Open Backlinks in Airefs. The list is pre-filtered to domains worth the effort: article-format pages only, at least 20 citations across our prompts, cited within the last 30 days, with major media and our own competitors excluded by default.
2. Work it in citation order. A mention on an already-trusted source can move answers almost immediately, which is exactly why this runs faster than publishing our own content.
3. Click a domain to load the outreach template, pre-filled with the specific articles already being cited. Pick the angle: link exchange, paid inclusion, or open-ended and let them choose.
4. Find a contact. The generic contact or support inbox is fine and often better than hunting the author, these sites mostly want to be reachable. Expect roughly a 1-in-10 reply rate, so work the list in volume rather than agonising over each one.
5. Send from our own inbox and personalise the opening line; the template is a starting point, not the email.
6. Most inclusions are paid. Get the number, then Human approval, the client approves any spend before we commit. Never commit their budget.
7. Mark emailed or dismissed so the domain moves to Archive and the active list stays honest. Follow up once after about a week.
8. A mention without a link still counts, the models read content, not the link graph. Do not turn down a no-follow placement.

Done when:

- the week's shortlist is emailed or archived, and

- any paid placement has written client approval.

Report the three layers in order of how much they can be trusted, and say out loud which is which.

1. Leads, the truth. Pull the "how did you hear about us?" answers from HubSpot for the period. Count leads and paying customers attributing discovery to ChatGPT, Perplexity or another assistant. This is the number that decides whether the channel is working.
2. Impressions, the leading indicator. Crawler activity moves before mention rate does. Report the trend by page and by LLM with "user searches only" on. Then read the pages table for the gap that matters: high impressions, no citation means the page is being read but is not authoritative enough, sharpen its structure and specificity rather than writing something new. No impressions at all means the models are not finding it, however much effort went in.
3. Share of voice, directional. Mention rate, rank and competitor comparison by tag. State the limitation plainly: it tracks a fixed prompt list with no volume data behind it, so it is a direction of travel, not market share.
4. Report correlation, not causation. Note which actions preceded which movement and build the picture over weeks. Never claim one article or one placement caused a change, the answers are probabilistic.
5. Promote what correlates with sustained lift; drop what does not. Retire dead prompts, expand the clusters that moved, and feed the priorities into next quarter's prompt set.
6. Human approval, sign off before anything client-facing goes out.

Done when:

- the report covers all three layers with their reliability stated, and

- next period's priorities are agreed.