I Run an Agency and Need to Show Clients How They Rank in AI Search Results

Clients will ask how they "rank" in AI search. Do not give them a Google average position with a new label. GEO rank is citation position inside the generated answer: named or not, linked or not, first among brands or fourth. Report that with screenshots and a frozen prompt list, or the QBR becomes an argument about vibes. Agencies that show proof keep the retainer. Agencies that show a mystery index do not.

Traditional rank tracking still belongs in the deck. It answers a different question. Mixing the two numbers is how a client concludes you "dropped from 3 to 7" when ChatGPT simply stopped naming them.

GEO rank is not Google rank

	Google rank	GEO "rank"
Unit	URL position on a SERP	Mention / citation slot in an answer
Winner	Ten blue links, plus features	Two to five named sources
Stability	Days to weeks	Hours to days
Proof	SERP snapshot	Answer screenshot + cited domains
Client translation	"We are #2 for this keyword"	"We are the second brand named, with a link, on this prompt"

Generative engine optimization is the discipline. The report is a prompt portfolio, not a keyword universe. If you only have Search Console, you can show Google generative impressions from the June 2026 reports. That is presence on Google's AI surfaces, not rank in ChatGPT. Bing's AI Performance covers Copilot. Everything else needs a tracker.

Promptwatch's Overview citation-type series is useful context in the appendix: in July 2026 listicles averaged about 18% of Overview citations (down from about 26% in Q1) and product pages about 16.3%, with product pages leading daily from July 28. Do not use the format mix as the client's rank. Use it to explain why citation now sits next to position in the same QBR.

What belongs in the monthly report

Keep the deck short enough that a CMO can retell it.

1. Scope slide. Exact prompt count, engines, and date range. If you changed prompts, say so. Silent prompt edits make every trend fake.

2. Scoreboard. For each money prompt: mentioned (Y/N), cited with URL (Y/N), position among named brands, sentiment or factual error flag. Roll up to share of voice versus a named competitor set — not versus "the internet."

3. Engine split. ChatGPT, Perplexity, Gemini, Claude, Copilot, AI Overviews as separate columns. A blended "AI rank" hides a Copilot win and a ChatGPT wipeout.

4. Proof appendix. One screenshot per notable win or loss. Clients believe pictures of the answer. They do not believe a 67 without a source.

5. Action list. Three URLs to fix, one third-party source to pursue, nothing else. If the tracker cannot name the competing URL, you are not ready to invoice for "insights."

Google's AI features docs are the right citation when a client asks for "the ranking factors." There is no extra technical checklist. AI optimization is still crawlable, useful, structured data matching visible text. Your job is to show whether that was enough this month.

Screenshots are the QBR, not decoration

AI answers are volatile and personalized. A number without a picture is how you get "that's not what I see on my laptop." Rules:

Capture the consumer-looking answer, not an API blob, when the client cares about what buyers see.
Store the date and the prompt on the image.
Show before/after on any claim you take credit for.
When the client is named without a link, show that too. "Mentioned" and "cited" are different slides.

This is why we weigh evidence in how we rank. After GSC and Bing, you still need a prompt-level screenshot log for ChatGPT and Perplexity. Profound will do that on an enterprise invoice. Otterly will do a thinner version. Promptwatch is the agency-shaped option in that same tracker set — daily six-engine runs, stored screenshots, alerts, competitor SOV, Agency at $199/mo with client workspaces and white-label reports, and no writer pretending to be the deliverable.

Packaging for different client types

SEO retainer clients. Add a two-page GEO insert behind the rank tracker. Same keywords where possible, rewritten as prompts. Do not replace the rank report in month one.

Content retainers. Lead with lost prompts and the competing URLs. That is the brief. If you also sell writing, separate the measurement fee from the production fee so you are not grading your own drafts with a vanity index.

PR / brand clients. Sentiment and wrong-fact flags matter more than citation position. A hallucinated lawsuit is a P0 even at 80% SOV.

Enterprise procurement. They will ask for methodology. Write it once: prompt source, refresh cadence, engines, how you handle logged-out vs logged-in, known limits. Point at our methodology if you want a public analog. Do not claim pixel-perfect reproduction of every personalized answer.

White-label and workspaces without lying

Clients should see their brand, not your 14 other logos. Minimum:

Separate workspaces per client.
Exports or scheduled PDFs that do not require a login you will regret sharing.
A URL you can put in the QBR appendix.
Alerts that can go to the client's Slack, not only yours.

If your tracker cannot do that, you will rebuild the report in Google Slides every month. That does not scale past five clients. The Agency tier exists for this. Pay for workspaces. Do not screenshot a shared dashboard and crop.

What not to put in the deck

A single sitewide "AI visibility score" with no prompt list. That number is a sales widget.
Traffic recovered as if citations caused it, unless you have a clean before/after and no other launch. Promptwatch's ChatGPT slot drop already showed citation seats can shrink (about 6.4 sources per response down to about 4.7–4.9 after GPT-5.3 on March 4, 2026) while classic rankings hold. Do not take credit for weather.
GPT-5 ranking-factor theater. OpenAI has not published one. If a client asks, say so and show OAI-SearchBot access plus live citation measurements.
Competitive scores against brands the client did not agree to track. Surprise enemies create surprise arguments.

A reporting cadence that survives contact with clients

Daily (you, not them): alerts on money-prompt losses. Fix or flag within 48 hours.
Monthly: scoreboard + three actions + screenshot appendix.
Quarterly: prompt-set review. Kill vanity prompts. Add the questions sales started hearing. Re-baseline so you do not compare a new set to an old one.

Measure for two weeks before you claim a baseline. We say the same in what GEO is. Agencies that "optimize" in week one are optimizing noise.

FAQ

Can I use average Google position as the AI rank?

No. A #1 URL can be uncited in the Overview. An unranked URL can be the source the model quotes. Report both. Do not average them.

Will Search Console be enough for client reporting?

It is enough for "Google showed you in an AI feature." It is not enough for ChatGPT, Perplexity, Claude, or prompt-level losses. Pair it with Bing, then a tracker.

How many prompts per client?

Start with 20–40 that map to revenue. A 400-prompt set looks diligent and produces a mushy SOV. Clients remember ten questions. Track those like keywords you would actually bid on.

What if the client's own ChatGPT session looks different?

Say so in the methodology slide. Personalized and logged-in answers diverge. Your job is a repeatable, dated sample — not to clone the CEO's chat history.

Should we white-label a writer as "AI rank optimization"?

No. Writing is a separate line item. Measurement with screenshots is what makes the rank claim believable.

What to do this week

Pick one client. Freeze 25 prompts with them on a call. Get written agreement on the competitor set.
Turn on GSC generative AI reports and Bing AI Performance on their properties. Screenshot the empty or early state for the appendix.
Run the 25 prompts once across six engines. Build the first scoreboard by hand if you have to.
Put those prompts on a daily tracker with screenshot storage and a client workspace. Compare Agency-tier options in the tools list.
Write the one-page methodology now, before the first QBR, so you are not inventing it under pressure.