Recommendations: the nightly project review
What the nightly review does
Every night a check that uses no AI reads what is already stored for each project and lists what is worth changing, each recommendation with the figures behind it. They are on the project page, in the Overview's To do list under Recommendations, and on the page each one is about: Keyword Audit (Worth a look), AI Visibility, the Completion hub, the AI Overview page, AI Settings and the services note; project settings show one line counting them. Each recommendation is a title and one sentence, with what was measured under Details. Nothing changes until you press a button, every button is the page's own, and nothing runs for a read-only account. The weekly email for each site also says, in the account edition, how many recommendations were raised or changed that week and which to look at first.
Reminders about what you measure
A run asks a fixed number of prompts. Your own questions come first, two are kept for questions about the business, and the rest go to your topics two at a time - so 10 prompts reach 4 topics, and a list of 42 topics is gone round about every 11 weeks on a weekly watchlist. When a list is gone round less often than about once a month, the review asks "Still asking 10 prompts over 42 topics?" and shows the trade both ways: how often each topic is reached now and at 25 prompts, how firm a rate from that many prompts can be, and what 25 would cost a month on your own AI key. Ask 25 prompts saves the watchlist at 25 and makes that count yours. Trim the list shows the topics already suggested for retiring: dropping a topic frees a turn for the others, it saves no calls. Keep as is is remembered; Show dismissed undoes it. A watchlist whose count you set yourself, or one with fixed questions, is never raised. Tracked prompts asked daily are listed with their answers a month both ways (daily, and weekly) and an estimate of each on your own AI key, plus what the engines reported where they report a cost. Ask them weekly moves each one to weekly; their answers so far are kept. Weekly measurement paused because your AI key failed three runs in a row is listed with the date of its last run. On the AI Overview page, a list that is full at daily checks shows what weekly would cost and how many searches the same budget would then cover; Check weekly changes the account's setting for every site. The topics suggested for retiring or measuring instead are listed beside the subjects no run asks, as one list.
Recommendations about your keys and models
If a provider has told us a model is retired, or OpenRouter no longer lists it, the review suggests a current model of the same maker - the cheapest one that searches the web, by the price list, with the others named beside it. If a key is out of credit or refused, it says what has stopped and sends you to the place to fix it: I've topped up, or AI Settings. If a model of the same maker, the same tier and the same or a newer generation costs less than the one AI Settings use for a provider (through OpenRouter, the one you use for each maker), it shows the saving on your own use of the last 30 days, measured where the provider reported the cost and estimated elsewhere. It never suggests a smaller or older model, and a model whose name does not say its tier and generation is not compared. There is one such recommendation per provider, listing every place the model is used. A watchlist that names an older model than the one AI Settings use for that provider gets a reminder of its own, on its project: "Still on gpt-5-mini? Your AI Settings use gpt-6.1-sol for OpenAI". Use the AI Settings model changes that watchlist's pick from its next run; Keep leaves it and is remembered. These recommendations are account-wide: a model on AI Settings is used for writing and by every watchlist that asks the provider without naming a model. Each names the projects it affects, and Change model or Switch asks you to confirm before it changes anything, with your key kept. Switch back undoes a Switch. A retired model is not switched back to. A same-maker change is a tick on the visibility trend, and the line carries on. The review also says when nothing asks any AI engine about a project, and when none of a watchlist's engines can see what you publish because none searches the web.
Crawlers and publishing
The same nightly review reads three things it already holds for the site: the latest site check, the AI crawler report, and what is waiting to be published or reviewed. It runs no check of its own, so after you change robots.txt, run the site check again and the recommendation closes on the next night. A crawler that fetches pages for an answer engine (OAI-SearchBot for ChatGPT search, PerplexityBot, Claude-SearchBot, Google-Extended for Gemini, Googlebot and Bingbot) is listed as Certain: while robots.txt disallows it, that engine cannot read the site to cite it. A training crawler (GPTBot, ClaudeBot) blocked is a lower note: answers can still cite you, and many sites choose it on purpose, so Dismiss keeps it from coming back. Crawler visits you cannot see are "not visible", never zero. Connect the site's Cloudflare zone (Growth and above) or install our WordPress plugin (every plan) to see them. It also says when IndexNow or Search Console needs attention. The Search Console AI setting is your answer, not our check: we cannot read it. If you said Exclude, the review repeats your answer; if you were not sure, or nobody has answered, it asks you to check it yourself. On the publishing side it lists subjects with evidence and no page on your site, pages going to another site, pages due for review, and Google files and pages held for your review for a week or more, with how many there are and the oldest's age. Each links to where the change is made; nothing is changed from the list.
Short topics, asked the way your customers ask
A topic of two or three words, such as "roof repair", leaves AI engines to guess where, for whom and why someone is asking. The review puts measured topics of one to three words that do not name your business together in one group on the project page, at most ten at a time. Rewrite each as a question a customer of yours might type - "Where can I get a quote for roof repair in Tampa?" - from the examples built from what you told us, or press Suggest with my AI to have your own AI key draft them: the line beside the button names the key, the model and about what it costs before you press. A suggestion is only a draft: nothing changes until you press Accept, and suggestions that name a business, drop the topic's words or add a place you have not told us are thrown away. Accept keeps the topic and its history: its counts include every run, in either wording, and the trend marks the day. A topic written as a question is asked word for word. Searches that need a short phrase (AI Overview, rank tracking, Reddit) keep using the short words. Keep as it is, and it is not suggested again; Show dismissed undoes that.
The AI review
The nightly check is free. When you press Review now, or once a week on Sunday night (UTC) if you switch "Weekly AI review" on, an AI review reads the recommendations. It may add a reading to a reminder (labelled "Our reading (AI review)", apart from what was measured), suggest measuring a subject that has open gaps in place of a topic the nightly check already suggests retiring, and put the three recommendations to do first at the top, one sentence each. It adds no figure and no change: an answer that names anything that was not measured, or any number that was not, is thrown away and nothing of it is kept. The review runs on your own AI key, on every plan and on the trial: on the model we list for that provider, or on your chosen model for a provider we do not list. Perplexity is never used: it searches on every call, and a review never searches. Each reading says which key it ran on and what it cost: measured where the provider reports it (OpenRouter), estimated from the list price otherwise. With no AI key, only the nightly checks run, and the Review panel says so. A check whose figures have not changed since it was last asked is not asked again, so a quiet project costs nothing. Review now says what it would run on and about what it costs before you press it. The weekly review is off until you switch it on, on the Review panel of any project or in AI Settings; the switch is for the whole account. Beside it is what a week costs at most on your key for all your projects - every check of every project asked afresh - so a quiet week costs less.
Filters, and many at once
A list of recommendations can be filtered by Certain or Worth a look, by what they are about (keywords, crawlers, models and so on) and by words in the title, and each filter shows how many it holds. Tick rows, or Select all shown, to Stop watching, Keep or put off many at once: what it does is said once before you confirm, and a row whose figures changed since you saw it is left alone and says so. The same words tracked in two campaigns show as one row that names both, and its buttons apply to both.